From 5a1110d499827131e8c05a82798f9bd1c9ac3abd Mon Sep 17 00:00:00 2001
From: Moses Paul R <iammosespaulr@gmail.com>
Date: Tue, 21 Jan 2025 10:28:55 +0000
Subject: [PATCH 01/46] add document provider (docx etc)

---
 convert.py                   |   2 +-
 marker/providers/document.py | 103 ++++++++
 marker/providers/registry.py |  31 ++-
 poetry.lock                  | 455 ++++++++++++++++++++++++++++++++++-
 pyproject.toml               |   2 +
 5 files changed, 587 insertions(+), 6 deletions(-)
 create mode 100644 marker/providers/document.py

diff --git a/convert.py b/convert.py
index 73e2a770..8a0505b1 100755
--- a/convert.py
+++ b/convert.py
@@ -1,4 +1,4 @@
 from marker.scripts import convert_cli
 
 if __name__ == "__main__":
-    convert_cli()
\ No newline at end of file
+    convert_cli()
diff --git a/marker/providers/document.py b/marker/providers/document.py
new file mode 100644
index 00000000..62abd4f6
--- /dev/null
+++ b/marker/providers/document.py
@@ -0,0 +1,103 @@
+import base64
+import logging
+import os
+import re
+from io import BytesIO
+
+import mammoth
+from PIL import Image
+from weasyprint import CSS, HTML
+
+from marker.providers.pdf import PdfProvider
+
+logging.getLogger('fontTools.subset').setLevel(logging.ERROR)
+logging.getLogger('fontTools.ttLib.ttFont').setLevel(logging.ERROR)
+
+css = '''
+    @page {
+        size: A4;
+        margin: 2cm;
+        @bottom-center {
+            content: counter(page);
+        }
+    }
+    
+    /* Force images to fit within page bounds */
+    img {
+        max-width: 100% !important;
+        max-height: 25cm !important;  /* A4 height minus margins */
+        object-fit: contain;
+        margin: 1em auto;
+    }
+    
+    /* Handle images that are inside centered paragraphs */
+    .center img {
+        margin-left: auto;
+        margin-right: auto;
+    }
+    
+    /* Prevent content overflow */
+    div, p, table {
+        max-width: 100%;
+        box-sizing: border-box;
+        overflow-wrap: break-word;
+    }
+'''
+
+
+class DocumentProvider(PdfProvider):
+    def __init__(self, filepath: str, config=None):
+        home_dir = os.path.expanduser("~")
+        rel_path = os.path.relpath(filepath, home_dir)
+        base_name, _ = os.path.splitext(rel_path)
+        self.temp_pdf_path = os.path.join('/tmp', f"{base_name}.pdf")
+
+        # Convert DOCX to PDF
+        try:
+            self.convert_docx_to_pdf(filepath)
+        except Exception as e:
+            raise RuntimeError(f"Failed to convert {filepath} to PDF: {e}")
+
+        # Initialize the PDF provider with the temp pdf path
+        super().__init__(self.temp_pdf_path, config)
+
+    def __del__(self):
+        if os.path.exists(self.temp_pdf_path):
+            print(f"Deleting temporary PDF file: {self.temp_pdf_path}")
+            os.remove(self.temp_pdf_path)
+
+    def convert_docx_to_pdf(self, filepath: str):
+        with open(filepath, "rb") as docx_file:
+            # we convert the docx to HTML
+            result = mammoth.convert_to_html(docx_file)
+            html = result.value
+
+            # We convert the HTML into a PDF
+            HTML(string=self._preprocess_base64_images(html)).write_pdf(
+                self.temp_pdf_path,
+                stylesheets=[CSS(string=css)]
+            )
+
+    @staticmethod
+    def _preprocess_base64_images(html_content):
+        pattern = r'data:([^;]+);base64,([^"\'>\s]+)'
+
+        def convert_image(match):
+            try:
+                full_data_uri = match.group(0)
+                base64_str = full_data_uri.split('base64,')[1]
+
+                img_data = base64.b64decode(base64_str)
+
+                with BytesIO(img_data) as bio:
+                    with Image.open(bio) as img:
+                        output = BytesIO()
+                        img.save(output, format=img.format)
+                        new_base64 = base64.b64encode(output.getvalue()).decode()
+                        return f'data:{match.group(1)};base64,{new_base64}'
+
+            except Exception as e:
+                print(e)
+                return ""  # we ditch broken images
+
+        return re.sub(pattern, convert_image, html_content)
diff --git a/marker/providers/registry.py b/marker/providers/registry.py
index 8f6d86fb..2d40c568 100644
--- a/marker/providers/registry.py
+++ b/marker/providers/registry.py
@@ -1,12 +1,37 @@
 import filetype
+import filetype.match as match
+from filetype.types import document
 
+from marker.providers.document import DocumentProvider
 from marker.providers.image import ImageProvider
 from marker.providers.pdf import PdfProvider
+from marker.providers.powerpoint import PowerPointProvider
+from marker.providers.spreadsheet import SpreadSheetProvider
 
 
 def provider_from_filepath(filepath: str):
-     kind = filetype.image_match(filepath)
-     if kind is not None:
+    if filetype.image_match(filepath) is not None:
         return ImageProvider
+    if match(
+            filepath, (
+                document.Doc(),
+                document.Docx(),
+                document.Odt()
+            )) is not None:
+        return DocumentProvider
+    if match(
+            filepath, (
+                document.Xls(),
+                document.Xlsx(),
+                document.Ods(),
+            )) is not None:
+        return SpreadSheetProvider
+    if match(
+            filepath, (
+                document.Ppt(),
+                document.Pptx(),
+                document.Odp(),
+            )) is not None:
+        return PowerPointProvider
 
-     return PdfProvider
\ No newline at end of file
+    return PdfProvider
diff --git a/poetry.lock b/poetry.lock
index 41e9e387..710da3c3 100644
--- a/poetry.lock
+++ b/poetry.lock
@@ -1,4 +1,4 @@
-# This file is automatically @generated by Poetry 1.8.2 and should not be changed by hand.
+# This file is automatically @generated by Poetry 1.8.5 and should not be changed by hand.
 
 [[package]]
 name = "aiohappyeyeballs"
@@ -401,6 +401,179 @@ files = [
     {file = "blinker-1.9.0.tar.gz", hash = "sha256:b4ce2265a7abece45e7cc896e98dbebe6cead56bcf805a3d23136d145f5445bf"},
 ]
 
+[[package]]
+name = "brotli"
+version = "1.1.0"
+description = "Python bindings for the Brotli compression library"
+optional = false
+python-versions = "*"
+files = [
+    {file = "Brotli-1.1.0-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:e1140c64812cb9b06c922e77f1c26a75ec5e3f0fb2bf92cc8c58720dec276752"},
+    {file = "Brotli-1.1.0-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:c8fd5270e906eef71d4a8d19b7c6a43760c6abcfcc10c9101d14eb2357418de9"},
+    {file = "Brotli-1.1.0-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:1ae56aca0402a0f9a3431cddda62ad71666ca9d4dc3a10a142b9dce2e3c0cda3"},
+    {file = "Brotli-1.1.0-cp310-cp310-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:43ce1b9935bfa1ede40028054d7f48b5469cd02733a365eec8a329ffd342915d"},
+    {file = "Brotli-1.1.0-cp310-cp310-manylinux_2_5_i686.manylinux1_i686.manylinux_2_12_i686.manylinux2010_i686.whl", hash = "sha256:7c4855522edb2e6ae7fdb58e07c3ba9111e7621a8956f481c68d5d979c93032e"},
+    {file = "Brotli-1.1.0-cp310-cp310-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_12_x86_64.manylinux2010_x86_64.whl", hash = "sha256:38025d9f30cf4634f8309c6874ef871b841eb3c347e90b0851f63d1ded5212da"},
+    {file = "Brotli-1.1.0-cp310-cp310-musllinux_1_1_aarch64.whl", hash = "sha256:e6a904cb26bfefc2f0a6f240bdf5233be78cd2488900a2f846f3c3ac8489ab80"},
+    {file = "Brotli-1.1.0-cp310-cp310-musllinux_1_1_i686.whl", hash = "sha256:a37b8f0391212d29b3a91a799c8e4a2855e0576911cdfb2515487e30e322253d"},
+    {file = "Brotli-1.1.0-cp310-cp310-musllinux_1_1_ppc64le.whl", hash = "sha256:e84799f09591700a4154154cab9787452925578841a94321d5ee8fb9a9a328f0"},
+    {file = "Brotli-1.1.0-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:f66b5337fa213f1da0d9000bc8dc0cb5b896b726eefd9c6046f699b169c41b9e"},
+    {file = "Brotli-1.1.0-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:5dab0844f2cf82be357a0eb11a9087f70c5430b2c241493fc122bb6f2bb0917c"},
+    {file = "Brotli-1.1.0-cp310-cp310-musllinux_1_2_i686.whl", hash = "sha256:e4fe605b917c70283db7dfe5ada75e04561479075761a0b3866c081d035b01c1"},
+    {file = "Brotli-1.1.0-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:1e9a65b5736232e7a7f91ff3d02277f11d339bf34099a56cdab6a8b3410a02b2"},
+    {file = "Brotli-1.1.0-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:58d4b711689366d4a03ac7957ab8c28890415e267f9b6589969e74b6e42225ec"},
+    {file = "Brotli-1.1.0-cp310-cp310-win32.whl", hash = "sha256:be36e3d172dc816333f33520154d708a2657ea63762ec16b62ece02ab5e4daf2"},
+    {file = "Brotli-1.1.0-cp310-cp310-win_amd64.whl", hash = "sha256:0c6244521dda65ea562d5a69b9a26120769b7a9fb3db2fe9545935ed6735b128"},
+    {file = "Brotli-1.1.0-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:a3daabb76a78f829cafc365531c972016e4aa8d5b4bf60660ad8ecee19df7ccc"},
+    {file = "Brotli-1.1.0-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:c8146669223164fc87a7e3de9f81e9423c67a79d6b3447994dfb9c95da16e2d6"},
+    {file = "Brotli-1.1.0-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:30924eb4c57903d5a7526b08ef4a584acc22ab1ffa085faceb521521d2de32dd"},
+    {file = "Brotli-1.1.0-cp311-cp311-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:ceb64bbc6eac5a140ca649003756940f8d6a7c444a68af170b3187623b43bebf"},
+    {file = "Brotli-1.1.0-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:a469274ad18dc0e4d316eefa616d1d0c2ff9da369af19fa6f3daa4f09671fd61"},
+    {file = "Brotli-1.1.0-cp311-cp311-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:524f35912131cc2cabb00edfd8d573b07f2d9f21fa824bd3fb19725a9cf06327"},
+    {file = "Brotli-1.1.0-cp311-cp311-musllinux_1_1_aarch64.whl", hash = "sha256:5b3cc074004d968722f51e550b41a27be656ec48f8afaeeb45ebf65b561481dd"},
+    {file = "Brotli-1.1.0-cp311-cp311-musllinux_1_1_i686.whl", hash = "sha256:19c116e796420b0cee3da1ccec3b764ed2952ccfcc298b55a10e5610ad7885f9"},
+    {file = "Brotli-1.1.0-cp311-cp311-musllinux_1_1_ppc64le.whl", hash = "sha256:510b5b1bfbe20e1a7b3baf5fed9e9451873559a976c1a78eebaa3b86c57b4265"},
+    {file = "Brotli-1.1.0-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:a1fd8a29719ccce974d523580987b7f8229aeace506952fa9ce1d53a033873c8"},
+    {file = "Brotli-1.1.0-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:c247dd99d39e0338a604f8c2b3bc7061d5c2e9e2ac7ba9cc1be5a69cb6cd832f"},
+    {file = "Brotli-1.1.0-cp311-cp311-musllinux_1_2_i686.whl", hash = "sha256:1b2c248cd517c222d89e74669a4adfa5577e06ab68771a529060cf5a156e9757"},
+    {file = "Brotli-1.1.0-cp311-cp311-musllinux_1_2_ppc64le.whl", hash = "sha256:2a24c50840d89ded6c9a8fdc7b6ed3692ed4e86f1c4a4a938e1e92def92933e0"},
+    {file = "Brotli-1.1.0-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:f31859074d57b4639318523d6ffdca586ace54271a73ad23ad021acd807eb14b"},
+    {file = "Brotli-1.1.0-cp311-cp311-win32.whl", hash = "sha256:39da8adedf6942d76dc3e46653e52df937a3c4d6d18fdc94a7c29d263b1f5b50"},
+    {file = "Brotli-1.1.0-cp311-cp311-win_amd64.whl", hash = "sha256:aac0411d20e345dc0920bdec5548e438e999ff68d77564d5e9463a7ca9d3e7b1"},
+    {file = "Brotli-1.1.0-cp312-cp312-macosx_10_13_universal2.whl", hash = "sha256:32d95b80260d79926f5fab3c41701dbb818fde1c9da590e77e571eefd14abe28"},
+    {file = "Brotli-1.1.0-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:b760c65308ff1e462f65d69c12e4ae085cff3b332d894637f6273a12a482d09f"},
+    {file = "Brotli-1.1.0-cp312-cp312-macosx_10_9_universal2.whl", hash = "sha256:316cc9b17edf613ac76b1f1f305d2a748f1b976b033b049a6ecdfd5612c70409"},
+    {file = "Brotli-1.1.0-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:caf9ee9a5775f3111642d33b86237b05808dafcd6268faa492250e9b78046eb2"},
+    {file = "Brotli-1.1.0-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:70051525001750221daa10907c77830bc889cb6d865cc0b813d9db7fefc21451"},
+    {file = "Brotli-1.1.0-cp312-cp312-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:7f4bf76817c14aa98cc6697ac02f3972cb8c3da93e9ef16b9c66573a68014f91"},
+    {file = "Brotli-1.1.0-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:d0c5516f0aed654134a2fc936325cc2e642f8a0e096d075209672eb321cff408"},
+    {file = "Brotli-1.1.0-cp312-cp312-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:6c3020404e0b5eefd7c9485ccf8393cfb75ec38ce75586e046573c9dc29967a0"},
+    {file = "Brotli-1.1.0-cp312-cp312-musllinux_1_1_aarch64.whl", hash = "sha256:4ed11165dd45ce798d99a136808a794a748d5dc38511303239d4e2363c0695dc"},
+    {file = "Brotli-1.1.0-cp312-cp312-musllinux_1_1_i686.whl", hash = "sha256:4093c631e96fdd49e0377a9c167bfd75b6d0bad2ace734c6eb20b348bc3ea180"},
+    {file = "Brotli-1.1.0-cp312-cp312-musllinux_1_1_ppc64le.whl", hash = "sha256:7e4c4629ddad63006efa0ef968c8e4751c5868ff0b1c5c40f76524e894c50248"},
+    {file = "Brotli-1.1.0-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:861bf317735688269936f755fa136a99d1ed526883859f86e41a5d43c61d8966"},
+    {file = "Brotli-1.1.0-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:87a3044c3a35055527ac75e419dfa9f4f3667a1e887ee80360589eb8c90aabb9"},
+    {file = "Brotli-1.1.0-cp312-cp312-musllinux_1_2_i686.whl", hash = "sha256:c5529b34c1c9d937168297f2c1fde7ebe9ebdd5e121297ff9c043bdb2ae3d6fb"},
+    {file = "Brotli-1.1.0-cp312-cp312-musllinux_1_2_ppc64le.whl", hash = "sha256:ca63e1890ede90b2e4454f9a65135a4d387a4585ff8282bb72964fab893f2111"},
+    {file = "Brotli-1.1.0-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:e79e6520141d792237c70bcd7a3b122d00f2613769ae0cb61c52e89fd3443839"},
+    {file = "Brotli-1.1.0-cp312-cp312-win32.whl", hash = "sha256:5f4d5ea15c9382135076d2fb28dde923352fe02951e66935a9efaac8f10e81b0"},
+    {file = "Brotli-1.1.0-cp312-cp312-win_amd64.whl", hash = "sha256:906bc3a79de8c4ae5b86d3d75a8b77e44404b0f4261714306e3ad248d8ab0951"},
+    {file = "Brotli-1.1.0-cp313-cp313-macosx_10_13_universal2.whl", hash = "sha256:8bf32b98b75c13ec7cf774164172683d6e7891088f6316e54425fde1efc276d5"},
+    {file = "Brotli-1.1.0-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:7bc37c4d6b87fb1017ea28c9508b36bbcb0c3d18b4260fcdf08b200c74a6aee8"},
+    {file = "Brotli-1.1.0-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:3c0ef38c7a7014ffac184db9e04debe495d317cc9c6fb10071f7fefd93100a4f"},
+    {file = "Brotli-1.1.0-cp313-cp313-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:91d7cc2a76b5567591d12c01f019dd7afce6ba8cba6571187e21e2fc418ae648"},
+    {file = "Brotli-1.1.0-cp313-cp313-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:a93dde851926f4f2678e704fadeb39e16c35d8baebd5252c9fd94ce8ce68c4a0"},
+    {file = "Brotli-1.1.0-cp313-cp313-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:f0db75f47be8b8abc8d9e31bc7aad0547ca26f24a54e6fd10231d623f183d089"},
+    {file = "Brotli-1.1.0-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:6967ced6730aed543b8673008b5a391c3b1076d834ca438bbd70635c73775368"},
+    {file = "Brotli-1.1.0-cp313-cp313-musllinux_1_2_i686.whl", hash = "sha256:7eedaa5d036d9336c95915035fb57422054014ebdeb6f3b42eac809928e40d0c"},
+    {file = "Brotli-1.1.0-cp313-cp313-musllinux_1_2_ppc64le.whl", hash = "sha256:d487f5432bf35b60ed625d7e1b448e2dc855422e87469e3f450aa5552b0eb284"},
+    {file = "Brotli-1.1.0-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:832436e59afb93e1836081a20f324cb185836c617659b07b129141a8426973c7"},
+    {file = "Brotli-1.1.0-cp313-cp313-win32.whl", hash = "sha256:43395e90523f9c23a3d5bdf004733246fba087f2948f87ab28015f12359ca6a0"},
+    {file = "Brotli-1.1.0-cp313-cp313-win_amd64.whl", hash = "sha256:9011560a466d2eb3f5a6e4929cf4a09be405c64154e12df0dd72713f6500e32b"},
+    {file = "Brotli-1.1.0-cp36-cp36m-macosx_10_9_x86_64.whl", hash = "sha256:a090ca607cbb6a34b0391776f0cb48062081f5f60ddcce5d11838e67a01928d1"},
+    {file = "Brotli-1.1.0-cp36-cp36m-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:2de9d02f5bda03d27ede52e8cfe7b865b066fa49258cbab568720aa5be80a47d"},
+    {file = "Brotli-1.1.0-cp36-cp36m-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:2333e30a5e00fe0fe55903c8832e08ee9c3b1382aacf4db26664a16528d51b4b"},
+    {file = "Brotli-1.1.0-cp36-cp36m-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:4d4a848d1837973bf0f4b5e54e3bec977d99be36a7895c61abb659301b02c112"},
+    {file = "Brotli-1.1.0-cp36-cp36m-manylinux_2_5_x86_64.manylinux1_x86_64.whl", hash = "sha256:fdc3ff3bfccdc6b9cc7c342c03aa2400683f0cb891d46e94b64a197910dc4064"},
+    {file = "Brotli-1.1.0-cp36-cp36m-musllinux_1_1_aarch64.whl", hash = "sha256:5eeb539606f18a0b232d4ba45adccde4125592f3f636a6182b4a8a436548b914"},
+    {file = "Brotli-1.1.0-cp36-cp36m-musllinux_1_1_i686.whl", hash = "sha256:fd5f17ff8f14003595ab414e45fce13d073e0762394f957182e69035c9f3d7c2"},
+    {file = "Brotli-1.1.0-cp36-cp36m-musllinux_1_1_ppc64le.whl", hash = "sha256:069a121ac97412d1fe506da790b3e69f52254b9df4eb665cd42460c837193354"},
+    {file = "Brotli-1.1.0-cp36-cp36m-musllinux_1_1_x86_64.whl", hash = "sha256:e93dfc1a1165e385cc8239fab7c036fb2cd8093728cbd85097b284d7b99249a2"},
+    {file = "Brotli-1.1.0-cp36-cp36m-musllinux_1_2_aarch64.whl", hash = "sha256:aea440a510e14e818e67bfc4027880e2fb500c2ccb20ab21c7a7c8b5b4703d75"},
+    {file = "Brotli-1.1.0-cp36-cp36m-musllinux_1_2_i686.whl", hash = "sha256:6974f52a02321b36847cd19d1b8e381bf39939c21efd6ee2fc13a28b0d99348c"},
+    {file = "Brotli-1.1.0-cp36-cp36m-musllinux_1_2_ppc64le.whl", hash = "sha256:a7e53012d2853a07a4a79c00643832161a910674a893d296c9f1259859a289d2"},
+    {file = "Brotli-1.1.0-cp36-cp36m-musllinux_1_2_x86_64.whl", hash = "sha256:d7702622a8b40c49bffb46e1e3ba2e81268d5c04a34f460978c6b5517a34dd52"},
+    {file = "Brotli-1.1.0-cp36-cp36m-win32.whl", hash = "sha256:a599669fd7c47233438a56936988a2478685e74854088ef5293802123b5b2460"},
+    {file = "Brotli-1.1.0-cp36-cp36m-win_amd64.whl", hash = "sha256:d143fd47fad1db3d7c27a1b1d66162e855b5d50a89666af46e1679c496e8e579"},
+    {file = "Brotli-1.1.0-cp37-cp37m-macosx_10_9_x86_64.whl", hash = "sha256:11d00ed0a83fa22d29bc6b64ef636c4552ebafcef57154b4ddd132f5638fbd1c"},
+    {file = "Brotli-1.1.0-cp37-cp37m-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:f733d788519c7e3e71f0855c96618720f5d3d60c3cb829d8bbb722dddce37985"},
+    {file = "Brotli-1.1.0-cp37-cp37m-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:929811df5462e182b13920da56c6e0284af407d1de637d8e536c5cd00a7daf60"},
+    {file = "Brotli-1.1.0-cp37-cp37m-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:0b63b949ff929fbc2d6d3ce0e924c9b93c9785d877a21a1b678877ffbbc4423a"},
+    {file = "Brotli-1.1.0-cp37-cp37m-manylinux_2_5_x86_64.manylinux1_x86_64.whl", hash = "sha256:d192f0f30804e55db0d0e0a35d83a9fead0e9a359a9ed0285dbacea60cc10a84"},
+    {file = "Brotli-1.1.0-cp37-cp37m-musllinux_1_1_aarch64.whl", hash = "sha256:f296c40e23065d0d6650c4aefe7470d2a25fffda489bcc3eb66083f3ac9f6643"},
+    {file = "Brotli-1.1.0-cp37-cp37m-musllinux_1_1_i686.whl", hash = "sha256:919e32f147ae93a09fe064d77d5ebf4e35502a8df75c29fb05788528e330fe74"},
+    {file = "Brotli-1.1.0-cp37-cp37m-musllinux_1_1_ppc64le.whl", hash = "sha256:23032ae55523cc7bccb4f6a0bf368cd25ad9bcdcc1990b64a647e7bbcce9cb5b"},
+    {file = "Brotli-1.1.0-cp37-cp37m-musllinux_1_1_x86_64.whl", hash = "sha256:224e57f6eac61cc449f498cc5f0e1725ba2071a3d4f48d5d9dffba42db196438"},
+    {file = "Brotli-1.1.0-cp37-cp37m-musllinux_1_2_aarch64.whl", hash = "sha256:cb1dac1770878ade83f2ccdf7d25e494f05c9165f5246b46a621cc849341dc01"},
+    {file = "Brotli-1.1.0-cp37-cp37m-musllinux_1_2_i686.whl", hash = "sha256:3ee8a80d67a4334482d9712b8e83ca6b1d9bc7e351931252ebef5d8f7335a547"},
+    {file = "Brotli-1.1.0-cp37-cp37m-musllinux_1_2_ppc64le.whl", hash = "sha256:5e55da2c8724191e5b557f8e18943b1b4839b8efc3ef60d65985bcf6f587dd38"},
+    {file = "Brotli-1.1.0-cp37-cp37m-musllinux_1_2_x86_64.whl", hash = "sha256:d342778ef319e1026af243ed0a07c97acf3bad33b9f29e7ae6a1f68fd083e90c"},
+    {file = "Brotli-1.1.0-cp37-cp37m-win32.whl", hash = "sha256:587ca6d3cef6e4e868102672d3bd9dc9698c309ba56d41c2b9c85bbb903cdb95"},
+    {file = "Brotli-1.1.0-cp37-cp37m-win_amd64.whl", hash = "sha256:2954c1c23f81c2eaf0b0717d9380bd348578a94161a65b3a2afc62c86467dd68"},
+    {file = "Brotli-1.1.0-cp38-cp38-macosx_10_9_universal2.whl", hash = "sha256:efa8b278894b14d6da122a72fefcebc28445f2d3f880ac59d46c90f4c13be9a3"},
+    {file = "Brotli-1.1.0-cp38-cp38-macosx_10_9_x86_64.whl", hash = "sha256:03d20af184290887bdea3f0f78c4f737d126c74dc2f3ccadf07e54ceca3bf208"},
+    {file = "Brotli-1.1.0-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:6172447e1b368dcbc458925e5ddaf9113477b0ed542df258d84fa28fc45ceea7"},
+    {file = "Brotli-1.1.0-cp38-cp38-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:a743e5a28af5f70f9c080380a5f908d4d21d40e8f0e0c8901604d15cfa9ba751"},
+    {file = "Brotli-1.1.0-cp38-cp38-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:0541e747cce78e24ea12d69176f6a7ddb690e62c425e01d31cc065e69ce55b48"},
+    {file = "Brotli-1.1.0-cp38-cp38-manylinux_2_5_x86_64.manylinux1_x86_64.whl", hash = "sha256:cdbc1fc1bc0bff1cef838eafe581b55bfbffaed4ed0318b724d0b71d4d377619"},
+    {file = "Brotli-1.1.0-cp38-cp38-musllinux_1_1_aarch64.whl", hash = "sha256:890b5a14ce214389b2cc36ce82f3093f96f4cc730c1cffdbefff77a7c71f2a97"},
+    {file = "Brotli-1.1.0-cp38-cp38-musllinux_1_1_i686.whl", hash = "sha256:1ab4fbee0b2d9098c74f3057b2bc055a8bd92ccf02f65944a241b4349229185a"},
+    {file = "Brotli-1.1.0-cp38-cp38-musllinux_1_1_ppc64le.whl", hash = "sha256:141bd4d93984070e097521ed07e2575b46f817d08f9fa42b16b9b5f27b5ac088"},
+    {file = "Brotli-1.1.0-cp38-cp38-musllinux_1_1_x86_64.whl", hash = "sha256:fce1473f3ccc4187f75b4690cfc922628aed4d3dd013d047f95a9b3919a86596"},
+    {file = "Brotli-1.1.0-cp38-cp38-musllinux_1_2_aarch64.whl", hash = "sha256:d2b35ca2c7f81d173d2fadc2f4f31e88cc5f7a39ae5b6db5513cf3383b0e0ec7"},
+    {file = "Brotli-1.1.0-cp38-cp38-musllinux_1_2_i686.whl", hash = "sha256:af6fa6817889314555aede9a919612b23739395ce767fe7fcbea9a80bf140fe5"},
+    {file = "Brotli-1.1.0-cp38-cp38-musllinux_1_2_ppc64le.whl", hash = "sha256:2feb1d960f760a575dbc5ab3b1c00504b24caaf6986e2dc2b01c09c87866a943"},
+    {file = "Brotli-1.1.0-cp38-cp38-musllinux_1_2_x86_64.whl", hash = "sha256:4410f84b33374409552ac9b6903507cdb31cd30d2501fc5ca13d18f73548444a"},
+    {file = "Brotli-1.1.0-cp38-cp38-win32.whl", hash = "sha256:db85ecf4e609a48f4b29055f1e144231b90edc90af7481aa731ba2d059226b1b"},
+    {file = "Brotli-1.1.0-cp38-cp38-win_amd64.whl", hash = "sha256:3d7954194c36e304e1523f55d7042c59dc53ec20dd4e9ea9d151f1b62b4415c0"},
+    {file = "Brotli-1.1.0-cp39-cp39-macosx_10_9_universal2.whl", hash = "sha256:5fb2ce4b8045c78ebbc7b8f3c15062e435d47e7393cc57c25115cfd49883747a"},
+    {file = "Brotli-1.1.0-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:7905193081db9bfa73b1219140b3d315831cbff0d8941f22da695832f0dd188f"},
+    {file = "Brotli-1.1.0-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:a77def80806c421b4b0af06f45d65a136e7ac0bdca3c09d9e2ea4e515367c7e9"},
+    {file = "Brotli-1.1.0-cp39-cp39-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:8dadd1314583ec0bf2d1379f7008ad627cd6336625d6679cf2f8e67081b83acf"},
+    {file = "Brotli-1.1.0-cp39-cp39-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:901032ff242d479a0efa956d853d16875d42157f98951c0230f69e69f9c09bac"},
+    {file = "Brotli-1.1.0-cp39-cp39-manylinux_2_5_x86_64.manylinux1_x86_64.whl", hash = "sha256:22fc2a8549ffe699bfba2256ab2ed0421a7b8fadff114a3d201794e45a9ff578"},
+    {file = "Brotli-1.1.0-cp39-cp39-musllinux_1_1_aarch64.whl", hash = "sha256:ae15b066e5ad21366600ebec29a7ccbc86812ed267e4b28e860b8ca16a2bc474"},
+    {file = "Brotli-1.1.0-cp39-cp39-musllinux_1_1_i686.whl", hash = "sha256:949f3b7c29912693cee0afcf09acd6ebc04c57af949d9bf77d6101ebb61e388c"},
+    {file = "Brotli-1.1.0-cp39-cp39-musllinux_1_1_ppc64le.whl", hash = "sha256:89f4988c7203739d48c6f806f1e87a1d96e0806d44f0fba61dba81392c9e474d"},
+    {file = "Brotli-1.1.0-cp39-cp39-musllinux_1_1_x86_64.whl", hash = "sha256:de6551e370ef19f8de1807d0a9aa2cdfdce2e85ce88b122fe9f6b2b076837e59"},
+    {file = "Brotli-1.1.0-cp39-cp39-musllinux_1_2_aarch64.whl", hash = "sha256:0737ddb3068957cf1b054899b0883830bb1fec522ec76b1098f9b6e0f02d9419"},
+    {file = "Brotli-1.1.0-cp39-cp39-musllinux_1_2_i686.whl", hash = "sha256:4f3607b129417e111e30637af1b56f24f7a49e64763253bbc275c75fa887d4b2"},
+    {file = "Brotli-1.1.0-cp39-cp39-musllinux_1_2_ppc64le.whl", hash = "sha256:6c6e0c425f22c1c719c42670d561ad682f7bfeeef918edea971a79ac5252437f"},
+    {file = "Brotli-1.1.0-cp39-cp39-musllinux_1_2_x86_64.whl", hash = "sha256:494994f807ba0b92092a163a0a283961369a65f6cbe01e8891132b7a320e61eb"},
+    {file = "Brotli-1.1.0-cp39-cp39-win32.whl", hash = "sha256:f0d8a7a6b5983c2496e364b969f0e526647a06b075d034f3297dc66f3b360c64"},
+    {file = "Brotli-1.1.0-cp39-cp39-win_amd64.whl", hash = "sha256:cdad5b9014d83ca68c25d2e9444e28e967ef16e80f6b436918c700c117a85467"},
+    {file = "Brotli-1.1.0.tar.gz", hash = "sha256:81de08ac11bcb85841e440c13611c00b67d3bf82698314928d0b676362546724"},
+]
+
+[[package]]
+name = "brotlicffi"
+version = "1.1.0.0"
+description = "Python CFFI bindings to the Brotli library"
+optional = false
+python-versions = ">=3.7"
+files = [
+    {file = "brotlicffi-1.1.0.0-cp37-abi3-macosx_10_9_x86_64.whl", hash = "sha256:9b7ae6bd1a3f0df532b6d67ff674099a96d22bc0948955cb338488c31bfb8851"},
+    {file = "brotlicffi-1.1.0.0-cp37-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:19ffc919fa4fc6ace69286e0a23b3789b4219058313cf9b45625016bf7ff996b"},
+    {file = "brotlicffi-1.1.0.0-cp37-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:9feb210d932ffe7798ee62e6145d3a757eb6233aa9a4e7db78dd3690d7755814"},
+    {file = "brotlicffi-1.1.0.0-cp37-abi3-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:84763dbdef5dd5c24b75597a77e1b30c66604725707565188ba54bab4f114820"},
+    {file = "brotlicffi-1.1.0.0-cp37-abi3-win32.whl", hash = "sha256:1b12b50e07c3911e1efa3a8971543e7648100713d4e0971b13631cce22c587eb"},
+    {file = "brotlicffi-1.1.0.0-cp37-abi3-win_amd64.whl", hash = "sha256:994a4f0681bb6c6c3b0925530a1926b7a189d878e6e5e38fae8efa47c5d9c613"},
+    {file = "brotlicffi-1.1.0.0-pp310-pypy310_pp73-macosx_10_9_x86_64.whl", hash = "sha256:2e4aeb0bd2540cb91b069dbdd54d458da8c4334ceaf2d25df2f4af576d6766ca"},
+    {file = "brotlicffi-1.1.0.0-pp310-pypy310_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:4b7b0033b0d37bb33009fb2fef73310e432e76f688af76c156b3594389d81391"},
+    {file = "brotlicffi-1.1.0.0-pp310-pypy310_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:54a07bb2374a1eba8ebb52b6fafffa2afd3c4df85ddd38fcc0511f2bb387c2a8"},
+    {file = "brotlicffi-1.1.0.0-pp310-pypy310_pp73-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:7901a7dc4b88f1c1475de59ae9be59799db1007b7d059817948d8e4f12e24e35"},
+    {file = "brotlicffi-1.1.0.0-pp310-pypy310_pp73-win_amd64.whl", hash = "sha256:ce01c7316aebc7fce59da734286148b1d1b9455f89cf2c8a4dfce7d41db55c2d"},
+    {file = "brotlicffi-1.1.0.0-pp37-pypy37_pp73-macosx_10_9_x86_64.whl", hash = "sha256:246f1d1a90279bb6069de3de8d75a8856e073b8ff0b09dcca18ccc14cec85979"},
+    {file = "brotlicffi-1.1.0.0-pp37-pypy37_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:cc4bc5d82bc56ebd8b514fb8350cfac4627d6b0743382e46d033976a5f80fab6"},
+    {file = "brotlicffi-1.1.0.0-pp37-pypy37_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:37c26ecb14386a44b118ce36e546ce307f4810bc9598a6e6cb4f7fca725ae7e6"},
+    {file = "brotlicffi-1.1.0.0-pp37-pypy37_pp73-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:ca72968ae4eaf6470498d5c2887073f7efe3b1e7d7ec8be11a06a79cc810e990"},
+    {file = "brotlicffi-1.1.0.0-pp37-pypy37_pp73-win_amd64.whl", hash = "sha256:add0de5b9ad9e9aa293c3aa4e9deb2b61e99ad6c1634e01d01d98c03e6a354cc"},
+    {file = "brotlicffi-1.1.0.0-pp38-pypy38_pp73-macosx_10_9_x86_64.whl", hash = "sha256:9b6068e0f3769992d6b622a1cd2e7835eae3cf8d9da123d7f51ca9c1e9c333e5"},
+    {file = "brotlicffi-1.1.0.0-pp38-pypy38_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:8557a8559509b61e65083f8782329188a250102372576093c88930c875a69838"},
+    {file = "brotlicffi-1.1.0.0-pp38-pypy38_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:2a7ae37e5d79c5bdfb5b4b99f2715a6035e6c5bf538c3746abc8e26694f92f33"},
+    {file = "brotlicffi-1.1.0.0-pp38-pypy38_pp73-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:391151ec86bb1c683835980f4816272a87eaddc46bb91cbf44f62228b84d8cca"},
+    {file = "brotlicffi-1.1.0.0-pp38-pypy38_pp73-win_amd64.whl", hash = "sha256:2f3711be9290f0453de8eed5275d93d286abe26b08ab4a35d7452caa1fef532f"},
+    {file = "brotlicffi-1.1.0.0-pp39-pypy39_pp73-macosx_10_9_x86_64.whl", hash = "sha256:1a807d760763e398bbf2c6394ae9da5815901aa93ee0a37bca5efe78d4ee3171"},
+    {file = "brotlicffi-1.1.0.0-pp39-pypy39_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:fa8ca0623b26c94fccc3a1fdd895be1743b838f3917300506d04aa3346fd2a14"},
+    {file = "brotlicffi-1.1.0.0-pp39-pypy39_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:3de0cf28a53a3238b252aca9fed1593e9d36c1d116748013339f0949bfc84112"},
+    {file = "brotlicffi-1.1.0.0-pp39-pypy39_pp73-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:6be5ec0e88a4925c91f3dea2bb0013b3a2accda6f77238f76a34a1ea532a1cb0"},
+    {file = "brotlicffi-1.1.0.0-pp39-pypy39_pp73-win_amd64.whl", hash = "sha256:d9eb71bb1085d996244439154387266fd23d6ad37161f6f52f1cd41dd95a3808"},
+    {file = "brotlicffi-1.1.0.0.tar.gz", hash = "sha256:b77827a689905143f87915310b93b273ab17888fd43ef350d4832c4a71083c13"},
+]
+
+[package.dependencies]
+cffi = ">=1.0.0"
+
 [[package]]
 name = "cachetools"
 version = "5.5.0"
@@ -617,6 +790,17 @@ files = [
 [package.dependencies]
 colorama = {version = "*", markers = "platform_system == \"Windows\""}
 
+[[package]]
+name = "cobble"
+version = "0.1.4"
+description = "Create data objects"
+optional = false
+python-versions = ">=3.5"
+files = [
+    {file = "cobble-0.1.4-py3-none-any.whl", hash = "sha256:36c91b1655e599fd428e2b95fdd5f0da1ca2e9f1abb0bc871dec21a0e78a2b44"},
+    {file = "cobble-0.1.4.tar.gz", hash = "sha256:de38be1539992c8a06e569630717c485a5f91be2192c461ea2b220607dfa78aa"},
+]
+
 [[package]]
 name = "colorama"
 version = "0.4.6"
@@ -645,6 +829,25 @@ traitlets = ">=4"
 [package.extras]
 test = ["pytest"]
 
+[[package]]
+name = "cssselect2"
+version = "0.7.0"
+description = "CSS selectors for Python ElementTree"
+optional = false
+python-versions = ">=3.7"
+files = [
+    {file = "cssselect2-0.7.0-py3-none-any.whl", hash = "sha256:fd23a65bfd444595913f02fc71f6b286c29261e354c41d722ca7a261a49b5969"},
+    {file = "cssselect2-0.7.0.tar.gz", hash = "sha256:1ccd984dab89fc68955043aca4e1b03e0cf29cad9880f6e28e3ba7a74b14aa5a"},
+]
+
+[package.dependencies]
+tinycss2 = "*"
+webencodings = "*"
+
+[package.extras]
+doc = ["sphinx", "sphinx_rtd_theme"]
+test = ["flake8", "isort", "pytest"]
+
 [[package]]
 name = "datasets"
 version = "2.21.0"
@@ -860,6 +1063,84 @@ files = [
     {file = "filetype-1.2.0.tar.gz", hash = "sha256:66b56cd6474bf41d8c54660347d37afcc3f7d1970648de365c102ef77548aadb"},
 ]
 
+[[package]]
+name = "fonttools"
+version = "4.55.3"
+description = "Tools to manipulate font files"
+optional = false
+python-versions = ">=3.8"
+files = [
+    {file = "fonttools-4.55.3-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:1dcc07934a2165ccdc3a5a608db56fb3c24b609658a5b340aee4ecf3ba679dc0"},
+    {file = "fonttools-4.55.3-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:f7d66c15ba875432a2d2fb419523f5d3d347f91f48f57b8b08a2dfc3c39b8a3f"},
+    {file = "fonttools-4.55.3-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:27e4ae3592e62eba83cd2c4ccd9462dcfa603ff78e09110680a5444c6925d841"},
+    {file = "fonttools-4.55.3-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:62d65a3022c35e404d19ca14f291c89cc5890032ff04f6c17af0bd1927299674"},
+    {file = "fonttools-4.55.3-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:d342e88764fb201286d185093781bf6628bbe380a913c24adf772d901baa8276"},
+    {file = "fonttools-4.55.3-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:dd68c87a2bfe37c5b33bcda0fba39b65a353876d3b9006fde3adae31f97b3ef5"},
+    {file = "fonttools-4.55.3-cp310-cp310-win32.whl", hash = "sha256:1bc7ad24ff98846282eef1cbeac05d013c2154f977a79886bb943015d2b1b261"},
+    {file = "fonttools-4.55.3-cp310-cp310-win_amd64.whl", hash = "sha256:b54baf65c52952db65df39fcd4820668d0ef4766c0ccdf32879b77f7c804d5c5"},
+    {file = "fonttools-4.55.3-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:8c4491699bad88efe95772543cd49870cf756b019ad56294f6498982408ab03e"},
+    {file = "fonttools-4.55.3-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:5323a22eabddf4b24f66d26894f1229261021dacd9d29e89f7872dd8c63f0b8b"},
+    {file = "fonttools-4.55.3-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:5480673f599ad410695ca2ddef2dfefe9df779a9a5cda89503881e503c9c7d90"},
+    {file = "fonttools-4.55.3-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:da9da6d65cd7aa6b0f806556f4985bcbf603bf0c5c590e61b43aa3e5a0f822d0"},
+    {file = "fonttools-4.55.3-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:e894b5bd60d9f473bed7a8f506515549cc194de08064d829464088d23097331b"},
+    {file = "fonttools-4.55.3-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:aee3b57643827e237ff6ec6d28d9ff9766bd8b21e08cd13bff479e13d4b14765"},
+    {file = "fonttools-4.55.3-cp311-cp311-win32.whl", hash = "sha256:eb6ca911c4c17eb51853143624d8dc87cdcdf12a711fc38bf5bd21521e79715f"},
+    {file = "fonttools-4.55.3-cp311-cp311-win_amd64.whl", hash = "sha256:6314bf82c54c53c71805318fcf6786d986461622dd926d92a465199ff54b1b72"},
+    {file = "fonttools-4.55.3-cp312-cp312-macosx_10_13_universal2.whl", hash = "sha256:f9e736f60f4911061235603a6119e72053073a12c6d7904011df2d8fad2c0e35"},
+    {file = "fonttools-4.55.3-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:7a8aa2c5e5b8b3bcb2e4538d929f6589a5c6bdb84fd16e2ed92649fb5454f11c"},
+    {file = "fonttools-4.55.3-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:07f8288aacf0a38d174445fc78377a97fb0b83cfe352a90c9d9c1400571963c7"},
+    {file = "fonttools-4.55.3-cp312-cp312-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:b8d5e8916c0970fbc0f6f1bece0063363bb5857a7f170121a4493e31c3db3314"},
+    {file = "fonttools-4.55.3-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:ae3b6600565b2d80b7c05acb8e24d2b26ac407b27a3f2e078229721ba5698427"},
+    {file = "fonttools-4.55.3-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:54153c49913f45065c8d9e6d0c101396725c5621c8aee744719300f79771d75a"},
+    {file = "fonttools-4.55.3-cp312-cp312-win32.whl", hash = "sha256:827e95fdbbd3e51f8b459af5ea10ecb4e30af50221ca103bea68218e9615de07"},
+    {file = "fonttools-4.55.3-cp312-cp312-win_amd64.whl", hash = "sha256:e6e8766eeeb2de759e862004aa11a9ea3d6f6d5ec710551a88b476192b64fd54"},
+    {file = "fonttools-4.55.3-cp313-cp313-macosx_10_13_universal2.whl", hash = "sha256:a430178ad3e650e695167cb53242dae3477b35c95bef6525b074d87493c4bf29"},
+    {file = "fonttools-4.55.3-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:529cef2ce91dc44f8e407cc567fae6e49a1786f2fefefa73a294704c415322a4"},
+    {file = "fonttools-4.55.3-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:8e75f12c82127486fac2d8bfbf5bf058202f54bf4f158d367e41647b972342ca"},
+    {file = "fonttools-4.55.3-cp313-cp313-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:859c358ebf41db18fb72342d3080bce67c02b39e86b9fbcf1610cca14984841b"},
+    {file = "fonttools-4.55.3-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:546565028e244a701f73df6d8dd6be489d01617863ec0c6a42fa25bf45d43048"},
+    {file = "fonttools-4.55.3-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:aca318b77f23523309eec4475d1fbbb00a6b133eb766a8bdc401faba91261abe"},
+    {file = "fonttools-4.55.3-cp313-cp313-win32.whl", hash = "sha256:8c5ec45428edaa7022f1c949a632a6f298edc7b481312fc7dc258921e9399628"},
+    {file = "fonttools-4.55.3-cp313-cp313-win_amd64.whl", hash = "sha256:11e5de1ee0d95af4ae23c1a138b184b7f06e0b6abacabf1d0db41c90b03d834b"},
+    {file = "fonttools-4.55.3-cp38-cp38-macosx_10_9_universal2.whl", hash = "sha256:caf8230f3e10f8f5d7593eb6d252a37caf58c480b19a17e250a63dad63834cf3"},
+    {file = "fonttools-4.55.3-cp38-cp38-macosx_10_9_x86_64.whl", hash = "sha256:b586ab5b15b6097f2fb71cafa3c98edfd0dba1ad8027229e7b1e204a58b0e09d"},
+    {file = "fonttools-4.55.3-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:a8c2794ded89399cc2169c4d0bf7941247b8d5932b2659e09834adfbb01589aa"},
+    {file = "fonttools-4.55.3-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:cf4fe7c124aa3f4e4c1940880156e13f2f4d98170d35c749e6b4f119a872551e"},
+    {file = "fonttools-4.55.3-cp38-cp38-musllinux_1_2_aarch64.whl", hash = "sha256:86721fbc389ef5cc1e2f477019e5069e8e4421e8d9576e9c26f840dbb04678de"},
+    {file = "fonttools-4.55.3-cp38-cp38-musllinux_1_2_x86_64.whl", hash = "sha256:89bdc5d88bdeec1b15af790810e267e8332d92561dce4f0748c2b95c9bdf3926"},
+    {file = "fonttools-4.55.3-cp38-cp38-win32.whl", hash = "sha256:bc5dbb4685e51235ef487e4bd501ddfc49be5aede5e40f4cefcccabc6e60fb4b"},
+    {file = "fonttools-4.55.3-cp38-cp38-win_amd64.whl", hash = "sha256:cd70de1a52a8ee2d1877b6293af8a2484ac82514f10b1c67c1c5762d38073e56"},
+    {file = "fonttools-4.55.3-cp39-cp39-macosx_10_9_universal2.whl", hash = "sha256:bdcc9f04b36c6c20978d3f060e5323a43f6222accc4e7fcbef3f428e216d96af"},
+    {file = "fonttools-4.55.3-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:c3ca99e0d460eff46e033cd3992a969658c3169ffcd533e0a39c63a38beb6831"},
+    {file = "fonttools-4.55.3-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:22f38464daa6cdb7b6aebd14ab06609328fe1e9705bb0fcc7d1e69de7109ee02"},
+    {file = "fonttools-4.55.3-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ed63959d00b61959b035c7d47f9313c2c1ece090ff63afea702fe86de00dbed4"},
+    {file = "fonttools-4.55.3-cp39-cp39-musllinux_1_2_aarch64.whl", hash = "sha256:5e8d657cd7326eeaba27de2740e847c6b39dde2f8d7cd7cc56f6aad404ddf0bd"},
+    {file = "fonttools-4.55.3-cp39-cp39-musllinux_1_2_x86_64.whl", hash = "sha256:fb594b5a99943042c702c550d5494bdd7577f6ef19b0bc73877c948a63184a32"},
+    {file = "fonttools-4.55.3-cp39-cp39-win32.whl", hash = "sha256:dc5294a3d5c84226e3dbba1b6f61d7ad813a8c0238fceea4e09aa04848c3d851"},
+    {file = "fonttools-4.55.3-cp39-cp39-win_amd64.whl", hash = "sha256:aedbeb1db64496d098e6be92b2e63b5fac4e53b1b92032dfc6988e1ea9134a4d"},
+    {file = "fonttools-4.55.3-py3-none-any.whl", hash = "sha256:f412604ccbeee81b091b420272841e5ec5ef68967a9790e80bffd0e30b8e2977"},
+    {file = "fonttools-4.55.3.tar.gz", hash = "sha256:3983313c2a04d6cc1fe9251f8fc647754cf49a61dac6cb1e7249ae67afaafc45"},
+]
+
+[package.dependencies]
+brotli = {version = ">=1.0.1", optional = true, markers = "platform_python_implementation == \"CPython\" and extra == \"woff\""}
+brotlicffi = {version = ">=0.8.0", optional = true, markers = "platform_python_implementation != \"CPython\" and extra == \"woff\""}
+zopfli = {version = ">=0.1.4", optional = true, markers = "extra == \"woff\""}
+
+[package.extras]
+all = ["brotli (>=1.0.1)", "brotlicffi (>=0.8.0)", "fs (>=2.2.0,<3)", "lxml (>=4.0)", "lz4 (>=1.7.4.2)", "matplotlib", "munkres", "pycairo", "scipy", "skia-pathops (>=0.5.0)", "sympy", "uharfbuzz (>=0.23.0)", "unicodedata2 (>=15.1.0)", "xattr", "zopfli (>=0.1.4)"]
+graphite = ["lz4 (>=1.7.4.2)"]
+interpolatable = ["munkres", "pycairo", "scipy"]
+lxml = ["lxml (>=4.0)"]
+pathops = ["skia-pathops (>=0.5.0)"]
+plot = ["matplotlib"]
+repacker = ["uharfbuzz (>=0.23.0)"]
+symfont = ["sympy"]
+type1 = ["xattr"]
+ufo = ["fs (>=2.2.0,<3)"]
+unicode = ["unicodedata2 (>=15.1.0)"]
+woff = ["brotli (>=1.0.1)", "brotlicffi (>=0.8.0)", "zopfli (>=0.1.4)"]
+
 [[package]]
 name = "fqdn"
 version = "1.5.1"
@@ -2056,6 +2337,20 @@ html5 = ["html5lib"]
 htmlsoup = ["BeautifulSoup4"]
 source = ["Cython (>=3.0.11)"]
 
+[[package]]
+name = "mammoth"
+version = "1.9.0"
+description = "Convert Word documents from docx to simple and clean HTML and Markdown"
+optional = false
+python-versions = ">=3.7"
+files = [
+    {file = "mammoth-1.9.0-py2.py3-none-any.whl", hash = "sha256:0eea277316586f0ca65d86834aec4de5a0572c83ec54b4991f9bb520a891150f"},
+    {file = "mammoth-1.9.0.tar.gz", hash = "sha256:74f5dae10ca240fd9b7a0e1a6deaebe0aad23bc590633ef6f5e868aa9b7042a6"},
+]
+
+[package.dependencies]
+cobble = ">=0.1.3,<0.2"
+
 [[package]]
 name = "markdown-it-py"
 version = "3.0.0"
@@ -3545,6 +3840,21 @@ numpy = ">=1.16.4"
 carto = ["pydeck-carto"]
 jupyter = ["ipykernel (>=5.1.2)", "ipython (>=5.8.0)", "ipywidgets (>=7,<8)", "traitlets (>=4.3.2)"]
 
+[[package]]
+name = "pydyf"
+version = "0.11.0"
+description = "A low-level PDF generator."
+optional = false
+python-versions = ">=3.8"
+files = [
+    {file = "pydyf-0.11.0-py3-none-any.whl", hash = "sha256:0aaf9e2ebbe786ec7a78ec3fbffa4cdcecde53fd6f563221d53c6bc1328848a3"},
+    {file = "pydyf-0.11.0.tar.gz", hash = "sha256:394dddf619cca9d0c55715e3c55ea121a9bf9cbc780cdc1201a2427917b86b64"},
+]
+
+[package.extras]
+doc = ["sphinx", "sphinx_rtd_theme"]
+test = ["pillow", "pytest", "ruff"]
+
 [[package]]
 name = "pygments"
 version = "2.18.0"
@@ -3595,6 +3905,21 @@ files = [
     {file = "pypdfium2-4.30.0.tar.gz", hash = "sha256:48b5b7e5566665bc1015b9d69c1ebabe21f6aee468b509531c3c8318eeee2e16"},
 ]
 
+[[package]]
+name = "pyphen"
+version = "0.17.2"
+description = "Pure Python module to hyphenate text"
+optional = false
+python-versions = ">=3.9"
+files = [
+    {file = "pyphen-0.17.2-py3-none-any.whl", hash = "sha256:3a07fb017cb2341e1d9ff31b8634efb1ae4dc4b130468c7c39dd3d32e7c3affd"},
+    {file = "pyphen-0.17.2.tar.gz", hash = "sha256:f60647a9c9b30ec6c59910097af82bc5dd2d36576b918e44148d8b07ef3b4aa3"},
+]
+
+[package.extras]
+doc = ["sphinx", "sphinx_rtd_theme"]
+test = ["pytest", "ruff"]
+
 [[package]]
 name = "pytest"
 version = "8.3.4"
@@ -4773,6 +5098,24 @@ webencodings = ">=0.4"
 doc = ["sphinx", "sphinx_rtd_theme"]
 test = ["pytest", "ruff"]
 
+[[package]]
+name = "tinyhtml5"
+version = "2.0.0"
+description = "HTML parser based on the WHATWG HTML specification"
+optional = false
+python-versions = ">=3.9"
+files = [
+    {file = "tinyhtml5-2.0.0-py3-none-any.whl", hash = "sha256:13683277c5b176d070f82d099d977194b7a1e26815b016114f581a74bbfbf47e"},
+    {file = "tinyhtml5-2.0.0.tar.gz", hash = "sha256:086f998833da24c300c414d9fe81d9b368fd04cb9d2596a008421cbc705fcfcc"},
+]
+
+[package.dependencies]
+webencodings = ">=0.5.1"
+
+[package.extras]
+doc = ["sphinx", "sphinx_rtd_theme"]
+test = ["pytest", "ruff"]
+
 [[package]]
 name = "tokenizers"
 version = "0.21.0"
@@ -5203,6 +5546,31 @@ files = [
     {file = "wcwidth-0.2.13.tar.gz", hash = "sha256:72ea0c06399eb286d978fdedb6923a9eb47e1c486ce63e9b4e64fc18303972b5"},
 ]
 
+[[package]]
+name = "weasyprint"
+version = "63.1"
+description = "The Awesome Document Factory"
+optional = false
+python-versions = ">=3.9"
+files = [
+    {file = "weasyprint-63.1-py3-none-any.whl", hash = "sha256:9d0319fe3ba553c9a77dc43a2d35b64a70c2b8809ad55a139a214803fde62bce"},
+    {file = "weasyprint-63.1.tar.gz", hash = "sha256:cb424e63e8dd3f14195bfe5f203527646aa40a2f00ac819f9d39b8304cec0044"},
+]
+
+[package.dependencies]
+cffi = ">=0.6"
+cssselect2 = ">=0.1"
+fonttools = {version = ">=4.0.0", extras = ["woff"]}
+Pillow = ">=9.1.0"
+pydyf = ">=0.11.0"
+Pyphen = ">=0.9.1"
+tinycss2 = ">=1.4.0"
+tinyhtml5 = ">=2.0.0b1"
+
+[package.extras]
+doc = ["sphinx", "sphinx_rtd_theme"]
+test = ["pytest", "ruff"]
+
 [[package]]
 name = "webcolors"
 version = "24.11.1"
@@ -5480,7 +5848,90 @@ idna = ">=2.0"
 multidict = ">=4.0"
 propcache = ">=0.2.0"
 
+[[package]]
+name = "zopfli"
+version = "0.2.3.post1"
+description = "Zopfli module for python"
+optional = false
+python-versions = ">=3.8"
+files = [
+    {file = "zopfli-0.2.3.post1-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:e0137dd64a493ba6a4be37405cfd6febe650a98cc1e9dca8f6b8c63b1db11b41"},
+    {file = "zopfli-0.2.3.post1-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:aa588b21044f8a74e423d8c8a4c7fc9988501878aacced793467010039c50734"},
+    {file = "zopfli-0.2.3.post1-cp310-cp310-manylinux_2_12_i686.manylinux2010_i686.whl", hash = "sha256:9f4a7ec2770e6af05f5a02733fd3900f30a9cd58e5d6d3727e14c5bcd6e7d587"},
+    {file = "zopfli-0.2.3.post1-cp310-cp310-manylinux_2_12_x86_64.manylinux2010_x86_64.whl", hash = "sha256:f7d69c1a7168ad0e9cb864e8663acb232986a0c9c9cb9801f56bf6214f53a54d"},
+    {file = "zopfli-0.2.3.post1-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:6c2d2bc8129707e34c51f9352c4636ca313b52350bbb7e04637c46c1818a2a70"},
+    {file = "zopfli-0.2.3.post1-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:39e576f93576c5c223b41d9c780bbb91fd6db4babf3223d2a4fe7bf568e2b5a8"},
+    {file = "zopfli-0.2.3.post1-cp310-cp310-musllinux_1_2_i686.whl", hash = "sha256:cbe6df25807227519debd1a57ab236f5f6bad441500e85b13903e51f93a43214"},
+    {file = "zopfli-0.2.3.post1-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:7cce242b5df12b2b172489daf19c32e5577dd2fac659eb4b17f6a6efb446fd5c"},
+    {file = "zopfli-0.2.3.post1-cp310-cp310-win32.whl", hash = "sha256:f815fcc2b2a457977724bad97fb4854022980f51ce7b136925e336b530545ae1"},
+    {file = "zopfli-0.2.3.post1-cp310-cp310-win_amd64.whl", hash = "sha256:0cc20b02a9531559945324c38302fd4ba763311632d0ec8a1a0aa9c10ea363e6"},
+    {file = "zopfli-0.2.3.post1-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:518f1f4ed35dd69ce06b552f84e6d081f07c552b4c661c5312d950a0b764a58a"},
+    {file = "zopfli-0.2.3.post1-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:615a8ac9dda265e9cc38b2a76c3142e4a9f30fea4a79c85f670850783bc6feb4"},
+    {file = "zopfli-0.2.3.post1-cp311-cp311-manylinux_2_12_i686.manylinux2010_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:a82fc2dbebe6eb908b9c665e71496f8525c1bc4d2e3a7a7722ef2b128b6227c8"},
+    {file = "zopfli-0.2.3.post1-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:37d011e92f7b9622742c905fdbed9920a1d0361df84142807ea2a528419dea7f"},
+    {file = "zopfli-0.2.3.post1-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:e63d558847166543c2c9789e6f985400a520b7eacc4b99181668b2c3aeadd352"},
+    {file = "zopfli-0.2.3.post1-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:60db20f06c3d4c5934b16cfa62a2cc5c3f0686bffe0071ed7804d3c31ab1a04e"},
+    {file = "zopfli-0.2.3.post1-cp311-cp311-musllinux_1_2_i686.whl", hash = "sha256:716cdbfc57bfd3d3e31a58e6246e8190e6849b7dbb7c4ce39ef8bbf0edb8f6d5"},
+    {file = "zopfli-0.2.3.post1-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:3a89277ed5f8c0fb2d0b46d669aa0633123aa7381f1f6118c12f15e0fb48f8ca"},
+    {file = "zopfli-0.2.3.post1-cp311-cp311-win32.whl", hash = "sha256:75a26a2307b10745a83b660c404416e984ee6fca515ec7f0765f69af3ce08072"},
+    {file = "zopfli-0.2.3.post1-cp311-cp311-win_amd64.whl", hash = "sha256:81c341d9bb87a6dbbb0d45d6e272aca80c7c97b4b210f9b6e233bf8b87242f29"},
+    {file = "zopfli-0.2.3.post1-cp312-cp312-macosx_10_13_universal2.whl", hash = "sha256:3f0197b6aa6eb3086ae9e66d6dd86c4d502b6c68b0ec490496348ae8c05ecaef"},
+    {file = "zopfli-0.2.3.post1-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:5fcfc0dc2761e4fcc15ad5d273b4d58c2e8e059d3214a7390d4d3c8e2aee644e"},
+    {file = "zopfli-0.2.3.post1-cp312-cp312-manylinux_2_12_i686.manylinux2010_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:cac2b37ab21c2b36a10b685b1893ebd6b0f83ae26004838ac817680881576567"},
+    {file = "zopfli-0.2.3.post1-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:8d5ab297d660b75c159190ce6d73035502310e40fd35170aed7d1a1aea7ddd65"},
+    {file = "zopfli-0.2.3.post1-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:9ba214f4f45bec195ee8559651154d3ac2932470b9d91c5715fc29c013349f8c"},
+    {file = "zopfli-0.2.3.post1-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:c1e0ed5d84ffa2d677cc9582fc01e61dab2e7ef8b8996e055f0a76167b1b94df"},
+    {file = "zopfli-0.2.3.post1-cp312-cp312-musllinux_1_2_i686.whl", hash = "sha256:bfa1eb759e07d8b7aa7a310a2bc535e127ee70addf90dc8d4b946b593c3e51a8"},
+    {file = "zopfli-0.2.3.post1-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:cd2c002f160502608dcc822ed2441a0f4509c52e86fcfd1a09e937278ed1ca14"},
+    {file = "zopfli-0.2.3.post1-cp312-cp312-win32.whl", hash = "sha256:7be5cc6732eb7b4df17305d8a7b293223f934a31783a874a01164703bc1be6cd"},
+    {file = "zopfli-0.2.3.post1-cp312-cp312-win_amd64.whl", hash = "sha256:4e50ffac74842c1c1018b9b73875a0d0a877c066ab06bf7cccbaa84af97e754f"},
+    {file = "zopfli-0.2.3.post1-cp313-cp313-macosx_10_13_universal2.whl", hash = "sha256:ecb7572df5372abce8073df078207d9d1749f20b8b136089916a4a0868d56051"},
+    {file = "zopfli-0.2.3.post1-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:a1cf720896d2ce998bc8e051d4b4ce0d8bec007aab6243102e8e1d22a0b2fb3f"},
+    {file = "zopfli-0.2.3.post1-cp313-cp313-manylinux_2_12_i686.manylinux2010_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:5aad740b4d4fcbaaae4887823925166ffd062db3b248b3f432198fc287381d1a"},
+    {file = "zopfli-0.2.3.post1-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:6617fb10f9e4393b331941861d73afb119cd847e88e4974bdbe8068ceef3f73f"},
+    {file = "zopfli-0.2.3.post1-cp313-cp313-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:a53b18797cdef27e019db595d66c4b077325afe2fd62145953275f53d84ce40c"},
+    {file = "zopfli-0.2.3.post1-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:b78008a69300d929ca2efeffec951b64a312e9a811e265ea4a907ab546d79fa6"},
+    {file = "zopfli-0.2.3.post1-cp313-cp313-musllinux_1_2_i686.whl", hash = "sha256:0aa5f90d6298bda02a95bc8dc8c3c19004d5a4e44bda00b67ca7431d857b4b54"},
+    {file = "zopfli-0.2.3.post1-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:2768c877f76c8a0e7519b1c86c93757f3c01492ddde55751e9988afb7eff64e1"},
+    {file = "zopfli-0.2.3.post1-cp313-cp313-win32.whl", hash = "sha256:71390dbd3fbf6ebea9a5d85ffed8c26ee1453ee09248e9b88486e30e0397b775"},
+    {file = "zopfli-0.2.3.post1-cp313-cp313-win_amd64.whl", hash = "sha256:a86eb88e06bd87e1fff31dac878965c26b0c26db59ddcf78bb0379a954b120de"},
+    {file = "zopfli-0.2.3.post1-cp38-cp38-macosx_10_9_universal2.whl", hash = "sha256:3827170de28faf144992d3d4dcf8f3998fe3c8a6a6f4a08f1d42c2ec6119d2bb"},
+    {file = "zopfli-0.2.3.post1-cp38-cp38-macosx_10_9_x86_64.whl", hash = "sha256:b0ec13f352ea5ae0fc91f98a48540512eed0767d0ec4f7f3cb92d92797983d18"},
+    {file = "zopfli-0.2.3.post1-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:5f272186e03ad55e7af09ab78055535c201b1a0bcc2944edb1768298d9c483a4"},
+    {file = "zopfli-0.2.3.post1-cp38-cp38-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:29ea74e72ffa6e291b8c6f2504ce6c146b4fe990c724c1450eb8e4c27fd31431"},
+    {file = "zopfli-0.2.3.post1-cp38-cp38-manylinux_2_5_x86_64.manylinux1_x86_64.whl", hash = "sha256:eb45a34f23da4f8bc712b6376ca5396914b0b7c09adbb001dad964eb7f3132f8"},
+    {file = "zopfli-0.2.3.post1-cp38-cp38-musllinux_1_2_aarch64.whl", hash = "sha256:6482db9876c68faac2d20a96b566ffbf65ddaadd97b222e4e73641f4f8722fc4"},
+    {file = "zopfli-0.2.3.post1-cp38-cp38-musllinux_1_2_i686.whl", hash = "sha256:95a260cafd56b8fffa679918937401c80bb38e1681c448b988022e4c3610965d"},
+    {file = "zopfli-0.2.3.post1-cp38-cp38-musllinux_1_2_x86_64.whl", hash = "sha256:676919fba7311125244eb0c4393679ac5fe856e5864a15d122bd815205369fa0"},
+    {file = "zopfli-0.2.3.post1-cp38-cp38-win32.whl", hash = "sha256:b9026a21b6d41eb0e2e63f5bc1242c3fcc43ecb770963cda99a4307863dac12e"},
+    {file = "zopfli-0.2.3.post1-cp38-cp38-win_amd64.whl", hash = "sha256:3c163911f8bad94b3e1db0a572e7c28ba681a0c91d0002ea1e4fa9264c21ef17"},
+    {file = "zopfli-0.2.3.post1-cp39-cp39-macosx_10_9_universal2.whl", hash = "sha256:b05296e8bc88c92e2b21e0a9bae4740c1551ee613c1d93a51fd28a7a0b2b6fbb"},
+    {file = "zopfli-0.2.3.post1-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:f12000a6accdd4bf0a3fa6eaa1b1c7a7bc80af0a2edf3f89d770d3dcce1d0e22"},
+    {file = "zopfli-0.2.3.post1-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:a241a68581d34d67b40c425cce3d1fd211c092f99d9250947824ccba9f491949"},
+    {file = "zopfli-0.2.3.post1-cp39-cp39-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:3657e416ffb8f31d9d3424af12122bb251befae109f2e271d87d825c92fc5b7b"},
+    {file = "zopfli-0.2.3.post1-cp39-cp39-manylinux_2_5_x86_64.manylinux1_x86_64.whl", hash = "sha256:4915a41375bdee4db749ecd07d985a0486eb688a6619f713b7bf6fbfd145e960"},
+    {file = "zopfli-0.2.3.post1-cp39-cp39-musllinux_1_2_aarch64.whl", hash = "sha256:bbe429fc50686bb2a2608a30843e36fbaa123462a5284f136c7d9e0145220bfd"},
+    {file = "zopfli-0.2.3.post1-cp39-cp39-musllinux_1_2_i686.whl", hash = "sha256:2345e713260a350bea0b01a816a469ea356bc2d63d009a0d777691ecbbcf7493"},
+    {file = "zopfli-0.2.3.post1-cp39-cp39-musllinux_1_2_x86_64.whl", hash = "sha256:fc39f5c27f962ec8660d8d20c24762431131b5d8c672b44b0a54cf2b5bcde9b9"},
+    {file = "zopfli-0.2.3.post1-cp39-cp39-win32.whl", hash = "sha256:9a6aec38a989bad7ddd1ef53f1265699e49e294d08231b5313d61293f3cd6237"},
+    {file = "zopfli-0.2.3.post1-cp39-cp39-win_amd64.whl", hash = "sha256:b3df42f52502438ee973042cc551877d24619fa1cd38ef7b7e9ac74200daca8b"},
+    {file = "zopfli-0.2.3.post1-pp310-pypy310_pp73-macosx_10_15_x86_64.whl", hash = "sha256:4c1226a7e2c7105ac31503a9bb97454743f55d88164d6d46bc138051b77f609b"},
+    {file = "zopfli-0.2.3.post1-pp310-pypy310_pp73-manylinux_2_12_i686.manylinux2010_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:48dba9251060289101343110ab47c0756f66f809bb4d1ddbb6d5c7e7752115c5"},
+    {file = "zopfli-0.2.3.post1-pp310-pypy310_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:89899641d4de97dbad8e0cde690040d078b6aea04066dacaab98e0b5a23573f2"},
+    {file = "zopfli-0.2.3.post1-pp310-pypy310_pp73-win_amd64.whl", hash = "sha256:3654bfc927bc478b1c3f3ff5056ed7b20a1a37fa108ca503256d0a699c03bbb1"},
+    {file = "zopfli-0.2.3.post1-pp38-pypy38_pp73-macosx_10_9_x86_64.whl", hash = "sha256:c4278d1873ce6e803e5d4f8d702fd3026bd67fca744aa98881324d1157ddf748"},
+    {file = "zopfli-0.2.3.post1-pp38-pypy38_pp73-win_amd64.whl", hash = "sha256:1d8cc06605519e82b16df090e17cb3990d1158861b2872c3117f1168777b81e4"},
+    {file = "zopfli-0.2.3.post1-pp39-pypy39_pp73-macosx_10_15_x86_64.whl", hash = "sha256:1f990634fd5c5c8ced8edddd8bd45fab565123b4194d6841e01811292650acae"},
+    {file = "zopfli-0.2.3.post1-pp39-pypy39_pp73-manylinux_2_12_i686.manylinux2010_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:91a2327a4d7e77471fa4fbb26991c6de4a738c6fc6a33e09bb25f56a870a4b7b"},
+    {file = "zopfli-0.2.3.post1-pp39-pypy39_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:8fbe5bcf10d01aab3513550f284c09fef32f342b36f56bfae2120a9c4d12c130"},
+    {file = "zopfli-0.2.3.post1-pp39-pypy39_pp73-win_amd64.whl", hash = "sha256:34a99592f3d9eb6f737616b5bd74b48a589fdb3cb59a01a50d636ea81d6af272"},
+    {file = "zopfli-0.2.3.post1.tar.gz", hash = "sha256:96484dc0f48be1c5d7ae9f38ed1ce41e3675fd506b27c11a6607f14b49101e99"},
+]
+
+[package.extras]
+test = ["pytest"]
+
 [metadata]
 lock-version = "2.0"
 python-versions = "^3.10"
-content-hash = "14d7c7764a6152887a987d8e1790e1e271d137ab88f5952e3043a45806334b6e"
+content-hash = "deba077d7c82f9718b904b70e0c86518c0fbe2321529b450cca0879f73ec4611"
diff --git a/pyproject.toml b/pyproject.toml
index 3152bd28..7744f376 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -36,6 +36,8 @@ google-generativeai = "^0.8.3"
 markdown2 = "^2.5.2"
 filetype = "^1.2.0"
 scikit-learn = "^1.6.1"
+mammoth = "^1.9.0"
+weasyprint = "^63.1"
 
 [tool.poetry.group.dev.dependencies]
 jupyter = "^1.0.0"

From f296cc4562160380ac363154d128cb70ce67ebf9 Mon Sep 17 00:00:00 2001
From: Moses Paul R <iammosespaulr@gmail.com>
Date: Wed, 22 Jan 2025 13:02:01 +0000
Subject: [PATCH 02/46] fix document styles, cleanup and add working
 spreadsheets

---
 marker/providers/document.py    |  69 ++++++++++---------
 marker/providers/spreadsheet.py | 117 ++++++++++++++++++++++++++++++++
 poetry.lock                     |  27 +++++++-
 pyproject.toml                  |   1 +
 4 files changed, 180 insertions(+), 34 deletions(-)
 create mode 100644 marker/providers/spreadsheet.py

diff --git a/marker/providers/document.py b/marker/providers/document.py
index 62abd4f6..025800d6 100644
--- a/marker/providers/document.py
+++ b/marker/providers/document.py
@@ -14,34 +14,40 @@
 logging.getLogger('fontTools.ttLib.ttFont').setLevel(logging.ERROR)
 
 css = '''
-    @page {
-        size: A4;
-        margin: 2cm;
-        @bottom-center {
-            content: counter(page);
-        }
-    }
-    
-    /* Force images to fit within page bounds */
-    img {
-        max-width: 100% !important;
-        max-height: 25cm !important;  /* A4 height minus margins */
-        object-fit: contain;
-        margin: 1em auto;
-    }
-    
-    /* Handle images that are inside centered paragraphs */
-    .center img {
-        margin-left: auto;
-        margin-right: auto;
-    }
-    
-    /* Prevent content overflow */
-    div, p, table {
-        max-width: 100%;
-        box-sizing: border-box;
-        overflow-wrap: break-word;
-    }
+@page {
+    size: A4;
+    margin: 2cm;
+}
+
+img {
+    max-width: 100%;
+    max-height: 25cm;
+    object-fit: contain;
+    margin: 12pt auto;
+}
+
+div, p {
+    max-width: 100%;
+    word-break: break-word;
+    font-size: 10pt;
+}
+
+table {
+    width: 100%;
+    border-collapse: collapse;
+    break-inside: auto;
+    font-size: 10pt;
+}
+
+tr {
+    break-inside: avoid;
+    page-break-inside: avoid;
+}
+
+td {
+    border: 0.75pt solid #000;
+    padding: 6pt;
+}
 '''
 
 
@@ -84,10 +90,7 @@ def _preprocess_base64_images(html_content):
 
         def convert_image(match):
             try:
-                full_data_uri = match.group(0)
-                base64_str = full_data_uri.split('base64,')[1]
-
-                img_data = base64.b64decode(base64_str)
+                img_data = base64.b64decode(match.group(2))
 
                 with BytesIO(img_data) as bio:
                     with Image.open(bio) as img:
@@ -98,6 +101,6 @@ def convert_image(match):
 
             except Exception as e:
                 print(e)
-                return ""  # we ditch broken images
+                return ""  # we ditch broken images as that breaks the PDF creation down the line
 
         return re.sub(pattern, convert_image, html_content)
diff --git a/marker/providers/spreadsheet.py b/marker/providers/spreadsheet.py
new file mode 100644
index 00000000..50ae30e2
--- /dev/null
+++ b/marker/providers/spreadsheet.py
@@ -0,0 +1,117 @@
+import os
+
+from openpyxl import load_workbook
+from openpyxl.worksheet.worksheet import Worksheet
+
+from marker.providers.pdf import PdfProvider
+from weasyprint import CSS, HTML
+
+css = '''
+@page {
+    size: A4 landscape;
+    margin: 1.5cm;
+}
+
+table {
+    width: 100%;
+    border-collapse: collapse;
+    break-inside: auto;
+    font-size: 10pt;
+}
+
+tr {
+    break-inside: avoid;
+    page-break-inside: avoid;
+}
+
+td {
+    border: 0.75pt solid #000;
+    padding: 6pt;
+}
+'''
+
+
+class SpreadSheetProvider(PdfProvider):
+    def __init__(self, filepath: str, config=None):
+        home_dir = os.path.expanduser("~")
+        rel_path = os.path.relpath(filepath, home_dir)
+        base_name, _ = os.path.splitext(rel_path)
+        self.temp_pdf_path = os.path.join('/tmp', f"{base_name}.pdf")
+
+        # Convert XLSX to PDF
+        try:
+            self.convert_xlsx_to_pdf(filepath)
+        except Exception as e:
+            raise RuntimeError(f"Failed to convert {filepath} to PDF: {e}")
+
+        # Initialize the PDF provider with the temp pdf path
+        super().__init__(self.temp_pdf_path, config)
+
+    def __del__(self):
+        if os.path.exists(self.temp_pdf_path):
+            print(f"Deleting temporary PDF file: {self.temp_pdf_path}")
+            os.remove(self.temp_pdf_path)
+
+    def convert_xlsx_to_pdf(self, filepath: str):
+        html = ""
+        workbook = load_workbook(filepath)
+        if workbook is not None:
+            for sheet_name in workbook.sheetnames:
+                sheet = workbook[sheet_name]
+                html += f'<div><h1>{sheet_name}</h1>' + self._excel_to_html_table(sheet) + '</div>'
+        else:
+            raise ValueError("Invalid XLSX file")
+
+        # We convert the HTML into a PDF
+        HTML(string=html).write_pdf(
+            self.temp_pdf_path,
+            stylesheets=[CSS(string=css)]
+        )
+
+    @staticmethod
+    def _get_merged_cell_ranges(sheet: Worksheet):
+        merged_info = {}
+        for merged_range in sheet.merged_cells.ranges:
+            min_col, min_row, max_col, max_row = merged_range.bounds
+            merged_info[(min_row, min_col)] = {
+                'rowspan': max_row - min_row + 1,
+                'colspan': max_col - min_col + 1,
+                'range': merged_range
+            }
+        return merged_info
+
+    def _excel_to_html_table(self, sheet: Worksheet):
+        merged_cells = self._get_merged_cell_ranges(sheet)
+
+        html = f'<table>'
+
+        # Track cells we should skip due to being part of a merge range
+        skip_cells = set()
+
+        for row_idx, row in enumerate(sheet.rows, 1):
+            html += '<tr>'
+            for col_idx, cell in enumerate(row, 1):
+                if (row_idx, col_idx) in skip_cells:
+                    continue
+
+                # Check if this cell is the start of a merged range
+                merge_info = merged_cells.get((row_idx, col_idx))
+                if merge_info:
+                    # Add cells to skip
+                    for r in range(row_idx, row_idx + merge_info['rowspan']):
+                        for c in range(col_idx, col_idx + merge_info['colspan']):
+                            if (r, c) != (row_idx, col_idx):
+                                skip_cells.add((r, c))
+
+                    # Add merged cell with rowspan/colspan
+                    value = cell.value if cell.value is not None else ''
+                    html += f'<td rowspan="{merge_info["rowspan"]}" colspan="{merge_info["colspan"]}">{value}'
+                else:
+                    # Regular cell
+                    value = cell.value if cell.value is not None else ''
+                    html += f'<td>{value}'
+
+                html += '</td>'
+            html += '</tr>'
+        html += '</table>'
+        return html
diff --git a/poetry.lock b/poetry.lock
index 710da3c3..4b25545c 100644
--- a/poetry.lock
+++ b/poetry.lock
@@ -974,6 +974,17 @@ files = [
     {file = "Distance-0.1.3.tar.gz", hash = "sha256:60807584f5b6003f5c521aa73f39f51f631de3be5cccc5a1d67166fcbf0d4551"},
 ]
 
+[[package]]
+name = "et-xmlfile"
+version = "2.0.0"
+description = "An implementation of lxml.xmlfile for the standard library"
+optional = false
+python-versions = ">=3.8"
+files = [
+    {file = "et_xmlfile-2.0.0-py3-none-any.whl", hash = "sha256:7a91720bc756843502c3b7504c77b8fe44217c85c537d85037f0f536151b2caa"},
+    {file = "et_xmlfile-2.0.0.tar.gz", hash = "sha256:dab3f4764309081ce75662649be815c4c9081e88f0837825f90fd28317d4da54"},
+]
+
 [[package]]
 name = "exceptiongroup"
 version = "1.2.2"
@@ -3076,6 +3087,20 @@ numpy = [
     {version = ">=1.26.0", markers = "python_version >= \"3.12\""},
 ]
 
+[[package]]
+name = "openpyxl"
+version = "3.1.5"
+description = "A Python library to read/write Excel 2010 xlsx/xlsm files"
+optional = false
+python-versions = ">=3.8"
+files = [
+    {file = "openpyxl-3.1.5-py2.py3-none-any.whl", hash = "sha256:5282c12b107bffeef825f4617dc029afaf41d0ea60823bbb665ef3079dc79de2"},
+    {file = "openpyxl-3.1.5.tar.gz", hash = "sha256:cf0e3cf56142039133628b5acffe8ef0c12bc902d2aadd3e0fe5878dc08d1050"},
+]
+
+[package.dependencies]
+et-xmlfile = "*"
+
 [[package]]
 name = "overrides"
 version = "7.7.0"
@@ -5934,4 +5959,4 @@ test = ["pytest"]
 [metadata]
 lock-version = "2.0"
 python-versions = "^3.10"
-content-hash = "deba077d7c82f9718b904b70e0c86518c0fbe2321529b450cca0879f73ec4611"
+content-hash = "42e3b1c26e61c9e61909cdd10ffa28b2218e265c3fb4d62f00466aa91434429b"
diff --git a/pyproject.toml b/pyproject.toml
index 7744f376..ba8e0ada 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -38,6 +38,7 @@ filetype = "^1.2.0"
 scikit-learn = "^1.6.1"
 mammoth = "^1.9.0"
 weasyprint = "^63.1"
+openpyxl = "^3.1.5"
 
 [tool.poetry.group.dev.dependencies]
 jupyter = "^1.0.0"

From d5a844a476038682c7376a5f48794a1bb1ee36f2 Mon Sep 17 00:00:00 2001
From: Moses Paul R <iammosespaulr@gmail.com>
Date: Thu, 23 Jan 2025 11:19:49 +0000
Subject: [PATCH 03/46] add powerpoint

---
 marker/providers/powerpoint.py | 252 +++++++++++++++++++++++++++++++++
 poetry.lock                    |  30 +++-
 pyproject.toml                 |   1 +
 3 files changed, 282 insertions(+), 1 deletion(-)
 create mode 100644 marker/providers/powerpoint.py

diff --git a/marker/providers/powerpoint.py b/marker/providers/powerpoint.py
new file mode 100644
index 00000000..8e3aea45
--- /dev/null
+++ b/marker/providers/powerpoint.py
@@ -0,0 +1,252 @@
+import base64
+import os
+import traceback
+
+from pptx import Presentation
+from pptx.enum.shapes import MSO_SHAPE_TYPE, PP_PLACEHOLDER
+from weasyprint import CSS, HTML
+
+from marker.providers.pdf import PdfProvider
+
+css = '''
+@page {
+    size: A4 landscape;
+    margin: 1.5cm;
+}
+
+table {
+    width: 100%;
+    border-collapse: collapse;
+    break-inside: auto;
+    font-size: 10pt;
+}
+
+tr {
+    break-inside: avoid;
+    page-break-inside: avoid;
+}
+
+td {
+    border: 0.75pt solid #000;
+    padding: 6pt;
+}
+
+img {
+    max-width: 100%;
+    height: auto;
+    object-fit: contain;
+}
+'''
+
+
+class PowerPointProvider(PdfProvider):
+    include_slide_number: bool = False
+
+    def __init__(self, filepath: str, config=None):
+        home_dir = os.path.expanduser("~")
+        rel_path = os.path.relpath(filepath, home_dir)
+        base_name, _ = os.path.splitext(rel_path)
+        self.temp_pdf_path = os.path.join('/tmp', f"{base_name}.pdf")
+
+        # Convert PPTX to PDF
+        try:
+            self.convert_pptx_to_pdf(filepath)
+        except Exception as e:
+            print(traceback.format_exc())
+            raise ValueError(f"Error converting PPTX to PDF: {e}")
+
+        # Initalize the PDF provider with the temp pdf path
+        super().__init__(self.temp_pdf_path, config)
+
+    def __del__(self):
+        if os.path.exists(self.temp_pdf_path):
+            print(f"Deleting temporary PDF file: {self.temp_pdf_path}")
+            # os.remove(self.temp_pdf_path)
+
+    def convert_pptx_to_pdf(self, filepath):
+        pptx = Presentation(filepath)
+
+        html_parts = []
+
+        for slide_index, slide in enumerate(pptx.slides):
+            html_parts.append(f"<section>")
+            if self.include_slide_number:
+                html_parts.append(f"<h2>Slide {slide_index + 1}</h2>")
+
+            # Process shapes in the slide
+            for shape in slide.shapes:
+                # If shape is a group shape, we recursively handle all grouped shapes
+                if shape.shape_type == MSO_SHAPE_TYPE.GROUP:
+                    html_parts.append(self._handle_group(shape))
+                    continue
+
+                # If shape is a table
+                if shape.has_table:
+                    html_parts.append(self._handle_table(shape))
+                    continue
+
+                # If shape is a picture
+                if shape.shape_type == MSO_SHAPE_TYPE.PICTURE:
+                    html_parts.append(self._handle_image(shape))
+                    continue
+
+                # If shape has text
+                if hasattr(shape, "text") and shape.text is not None:
+                    if shape.has_text_frame:
+                        # Distinguish placeholders (title, subtitle, etc.)
+                        html_parts.append(self._handle_text(shape))
+                    else:
+                        html_parts.append(f"<p>{self._escape_html(shape.text)}</p>")
+
+            html_parts.append(f"</section>")
+
+        html = '\n'.join(html_parts)
+
+        # We convert the HTML into a PDF
+        open(self.temp_pdf_path + '.html', "w").write(html)
+        print(self.temp_pdf_path + '.html')
+
+        HTML(string=html).write_pdf(
+            self.temp_pdf_path,
+            stylesheets=[CSS(string=css)]
+        )
+
+    def _handle_group(self, group_shape) -> str:
+        """
+        Recursively handle shapes in a group. Returns HTML string for the entire group.
+        """
+
+        group_parts = []
+        for shape in group_shape.shapes:
+            if shape.shape_type == MSO_SHAPE_TYPE.GROUP:
+                group_parts.append(self._handle_group(shape))
+                continue
+
+            if shape.has_table:
+                group_parts.append(self._handle_table(shape))
+                continue
+
+            if shape.shape_type == MSO_SHAPE_TYPE.PICTURE:
+                group_parts.append(self._handle_image(shape))
+                continue
+
+            if hasattr(shape, "text"):
+                if shape.has_text_frame:
+                    group_parts.append(self._handle_text(shape))
+                else:
+                    group_parts.append(f"<p>{self._escape_html(shape.text)}</p>")
+
+        return "".join(group_parts)
+
+    def _handle_text(self, shape) -> str:
+        """
+        Processes shape text, including bullet/numbered list detection and placeholders
+        (title, subtitle, etc.). Returns HTML for the text block(s).
+        """
+
+        # Distinguish placeholders to see if it's a title or subtitle
+        label_html_tag = "p"
+        if shape.is_placeholder:
+            placeholder_type = shape.placeholder_format.type
+            if placeholder_type in [PP_PLACEHOLDER.TITLE, PP_PLACEHOLDER.CENTER_TITLE]:
+                label_html_tag = "h3"
+            elif placeholder_type == PP_PLACEHOLDER.SUBTITLE:
+                label_html_tag = "h4"
+
+        # Keep track of whether we are currently in a <ul> or <ol>
+        html_parts = []
+        list_open = False
+        list_type = None  # "ul" or "ol"
+
+        for paragraph in shape.text_frame.paragraphs:
+            p_el = paragraph._element
+            # Check bullet
+            bullet_char = p_el.find(".//a:buChar", namespaces=p_el.nsmap)
+            bullet_num = p_el.find(".//a:buAutoNum", namespaces=p_el.nsmap)
+
+            is_bullet = (bullet_char is not None) or (paragraph.level > 0)
+            is_numbered = (bullet_num is not None)
+
+            # If the paragraph is bullet or numbered
+            if is_bullet or is_numbered:
+                # Decide if we need to start a new list or continue an existing one
+                current_list_type = "ol" if is_numbered else "ul"
+                if not list_open:
+                    # Start new
+                    list_open = True
+                    list_type = current_list_type
+                    html_parts.append(f"<{list_type}>")
+
+                elif list_open and list_type != current_list_type:
+                    # Close old list, start new
+                    html_parts.append(f"</{list_type}>")
+                    list_type = current_list_type
+                    html_parts.append(f"<{list_type}>")
+
+                # Build the bullet (li) text from all runs in the paragraph
+                p_text = "".join(run.text for run in paragraph.runs)
+                if p_text:
+                    html_parts.append(f"<li>{self._escape_html(p_text)}</li>")
+
+            else:
+                # If we were in a list, we need to close it
+                if list_open:
+                    html_parts.append(f"</{list_type}>")
+                    list_open = False
+                    list_type = None
+
+                # Now it's just a normal paragraph
+                # Gather the paragraph text from runs
+                p_text = "".join(run.text for run in paragraph.runs)
+                if p_text:
+                    # If we know it's a slide title, we can use <h3> or so
+                    html_parts.append(f"<{label_html_tag}>{self._escape_html(p_text)}</{label_html_tag}>")
+
+        # If the text frame ended and we still have an open list, close it
+        if list_open:
+            html_parts.append(f"</{list_type}>")
+
+        return "".join(html_parts)
+
+    def _handle_image(self, shape) -> str:
+        """
+        Embeds the image as a base64 <img> in HTML.
+        """
+        image = shape.image
+        image_bytes = image.blob
+
+        try:
+            img_str = base64.b64encode(image_bytes).decode('utf-8')
+            return f"<img src='data:{image.content_type};base64,{img_str}' />"
+        except Exception as e:
+            print(f"Warning: image cannot be loaded by Pillow: {e}")
+            return ""
+
+    def _handle_table(self, shape) -> str:
+        """
+        Renders a shape's table as an HTML <table>.
+        """
+        table_html = []
+        table_html.append("<table border='1'>")
+
+        for row in shape.table.rows:
+            row_html = ["<tr>"]
+            for cell in row.cells:
+                row_html.append(f"<td>{self._escape_html(cell.text)}</td>")
+            row_html.append("</tr>")
+            table_html.append("".join(row_html))
+
+        table_html.append("</table>")
+        return "".join(table_html)
+
+    def _escape_html(self, text: str) -> str:
+        """
+        Minimal escaping for HTML special characters.
+        """
+        return (
+            text.replace("&", "&amp;")
+                .replace("<", "&lt;")
+                .replace(">", "&gt;")
+                .replace('"', "&quot;")
+                .replace("'", "&#39;")
+        )
diff --git a/poetry.lock b/poetry.lock
index 4b25545c..f9ceaf6a 100644
--- a/poetry.lock
+++ b/poetry.lock
@@ -4037,6 +4037,23 @@ files = [
     {file = "python_multipart-0.0.16.tar.gz", hash = "sha256:8dee37b88dab9b59922ca173c35acb627cc12ec74019f5cd4578369c6df36554"},
 ]
 
+[[package]]
+name = "python-pptx"
+version = "1.0.2"
+description = "Create, read, and update PowerPoint 2007+ (.pptx) files."
+optional = false
+python-versions = ">=3.8"
+files = [
+    {file = "python_pptx-1.0.2-py3-none-any.whl", hash = "sha256:160838e0b8565a8b1f67947675886e9fea18aa5e795db7ae531606d68e785cba"},
+    {file = "python_pptx-1.0.2.tar.gz", hash = "sha256:479a8af0eaf0f0d76b6f00b0887732874ad2e3188230315290cd1f9dd9cc7095"},
+]
+
+[package.dependencies]
+lxml = ">=3.1.0"
+Pillow = ">=3.3.2"
+typing-extensions = ">=4.9.0"
+XlsxWriter = ">=0.5.7"
+
 [[package]]
 name = "pytz"
 version = "2024.2"
@@ -5645,6 +5662,17 @@ files = [
     {file = "widgetsnbextension-4.0.13.tar.gz", hash = "sha256:ffcb67bc9febd10234a362795f643927f4e0c05d9342c727b65d2384f8feacb6"},
 ]
 
+[[package]]
+name = "xlsxwriter"
+version = "3.2.1"
+description = "A Python module for creating Excel XLSX files."
+optional = false
+python-versions = ">=3.6"
+files = [
+    {file = "XlsxWriter-3.2.1-py3-none-any.whl", hash = "sha256:7e8f7c60b7a1660ef791d46ab5de78469cb978b991ca841af61f5832d2f9f4fe"},
+    {file = "XlsxWriter-3.2.1.tar.gz", hash = "sha256:97618759cb264fb6a93397f660cca156ffa9561743b1823dafb60dc4474e1902"},
+]
+
 [[package]]
 name = "xxhash"
 version = "3.5.0"
@@ -5959,4 +5987,4 @@ test = ["pytest"]
 [metadata]
 lock-version = "2.0"
 python-versions = "^3.10"
-content-hash = "42e3b1c26e61c9e61909cdd10ffa28b2218e265c3fb4d62f00466aa91434429b"
+content-hash = "429b563e9a609f51ba8185407ef5ef1219caf582b09386f5dca4740ff4386ada"
diff --git a/pyproject.toml b/pyproject.toml
index ba8e0ada..c1d6c56c 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -39,6 +39,7 @@ scikit-learn = "^1.6.1"
 mammoth = "^1.9.0"
 weasyprint = "^63.1"
 openpyxl = "^3.1.5"
+python-pptx = "^1.0.2"
 
 [tool.poetry.group.dev.dependencies]
 jupyter = "^1.0.0"

From 5795994395b239f8722520e43765669769524ded Mon Sep 17 00:00:00 2001
From: Moses Paul R <iammosespaulr@gmail.com>
Date: Thu, 23 Jan 2025 11:21:08 +0000
Subject: [PATCH 04/46] fix powerpoint temp file

---
 marker/providers/powerpoint.py | 5 +----
 1 file changed, 1 insertion(+), 4 deletions(-)

diff --git a/marker/providers/powerpoint.py b/marker/providers/powerpoint.py
index 8e3aea45..433f7f4b 100644
--- a/marker/providers/powerpoint.py
+++ b/marker/providers/powerpoint.py
@@ -61,7 +61,7 @@ def __init__(self, filepath: str, config=None):
     def __del__(self):
         if os.path.exists(self.temp_pdf_path):
             print(f"Deleting temporary PDF file: {self.temp_pdf_path}")
-            # os.remove(self.temp_pdf_path)
+            os.remove(self.temp_pdf_path)
 
     def convert_pptx_to_pdf(self, filepath):
         pptx = Presentation(filepath)
@@ -103,9 +103,6 @@ def convert_pptx_to_pdf(self, filepath):
         html = '\n'.join(html_parts)
 
         # We convert the HTML into a PDF
-        open(self.temp_pdf_path + '.html', "w").write(html)
-        print(self.temp_pdf_path + '.html')
-
         HTML(string=html).write_pdf(
             self.temp_pdf_path,
             stylesheets=[CSS(string=css)]

From c1182e679f470a784db9b1c934ebbd14e9c4d81c Mon Sep 17 00:00:00 2001
From: Moses Paul R <iammosespaulr@gmail.com>
Date: Thu, 23 Jan 2025 11:27:26 +0000
Subject: [PATCH 05/46] update registry [skip ci]

---
 marker/providers/registry.py | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/marker/providers/registry.py b/marker/providers/registry.py
index 2d40c568..f3dfb812 100644
--- a/marker/providers/registry.py
+++ b/marker/providers/registry.py
@@ -1,6 +1,6 @@
 import filetype
 import filetype.match as match
-from filetype.types import document
+from filetype.types import archive, document
 
 from marker.providers.document import DocumentProvider
 from marker.providers.image import ImageProvider
@@ -12,6 +12,8 @@
 def provider_from_filepath(filepath: str):
     if filetype.image_match(filepath) is not None:
         return ImageProvider
+    if match(filepath, (archive.Pdf(),)) is not None:
+        return PdfProvider
     if match(
             filepath, (
                 document.Doc(),

From f29c22555c7f1d0fc5c8f0f9f8b8c091f1aeea4e Mon Sep 17 00:00:00 2001
From: Moses Paul R <iammosespaulr@gmail.com>
Date: Wed, 5 Feb 2025 11:21:43 +0000
Subject: [PATCH 06/46] update poetry lock [skip ci]

---
 poetry.lock | 173 ++++++++++++++++++++++++++--------------------------
 1 file changed, 88 insertions(+), 85 deletions(-)

diff --git a/poetry.lock b/poetry.lock
index 38831bf7..ca581c10 100644
--- a/poetry.lock
+++ b/poetry.lock
@@ -339,31 +339,32 @@ tests-mypy = ["mypy (>=1.11.1)", "pytest-mypy-plugins"]
 
 [[package]]
 name = "babel"
-version = "2.16.0"
+version = "2.17.0"
 description = "Internationalization utilities"
 optional = false
 python-versions = ">=3.8"
 files = [
-    {file = "babel-2.16.0-py3-none-any.whl", hash = "sha256:368b5b98b37c06b7daf6696391c3240c938b37767d4584413e8438c5c435fa8b"},
-    {file = "babel-2.16.0.tar.gz", hash = "sha256:d1f3554ca26605fe173f3de0c65f750f5a42f924499bf134de6423582298e316"},
+    {file = "babel-2.17.0-py3-none-any.whl", hash = "sha256:4d0b53093fdfb4b21c92b5213dba5a1b23885afa8383709427046b21c366e5f2"},
+    {file = "babel-2.17.0.tar.gz", hash = "sha256:0c54cffb19f690cdcc52a3b50bcbf71e07a808d1c80d549f2459b9d2cf0afb9d"},
 ]
 
 [package.extras]
-dev = ["freezegun (>=1.0,<2.0)", "pytest (>=6.0)", "pytest-cov"]
+dev = ["backports.zoneinfo", "freezegun (>=1.0,<2.0)", "jinja2 (>=3.0)", "pytest (>=6.0)", "pytest-cov", "pytz", "setuptools", "tzdata"]
 
 [[package]]
 name = "beautifulsoup4"
-version = "4.12.3"
+version = "4.13.3"
 description = "Screen-scraping library"
 optional = false
-python-versions = ">=3.6.0"
+python-versions = ">=3.7.0"
 files = [
-    {file = "beautifulsoup4-4.12.3-py3-none-any.whl", hash = "sha256:b80878c9f40111313e55da8ba20bdba06d8fa3969fc68304167741bbf9e082ed"},
-    {file = "beautifulsoup4-4.12.3.tar.gz", hash = "sha256:74e3d1928edc070d21748185c46e3fb33490f22f52a3addee9aee0f4f7781051"},
+    {file = "beautifulsoup4-4.13.3-py3-none-any.whl", hash = "sha256:99045d7d3f08f91f0d656bc9b7efbae189426cd913d830294a15eefa0ea4df16"},
+    {file = "beautifulsoup4-4.13.3.tar.gz", hash = "sha256:1bd32405dacc920b42b83ba01644747ed77456a65760e285fbc47633ceddaf8b"},
 ]
 
 [package.dependencies]
 soupsieve = ">1.2"
+typing-extensions = ">=4.0.0"
 
 [package.extras]
 cchardet = ["cchardet"]
@@ -1076,61 +1077,61 @@ files = [
 
 [[package]]
 name = "fonttools"
-version = "4.55.4"
+version = "4.55.8"
 description = "Tools to manipulate font files"
 optional = false
 python-versions = ">=3.8"
 files = [
-    {file = "fonttools-4.55.4-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:3b332ea7b7f5f3d99f9bc5a28a23c3824ae72711abf7c4e1d62fa21699fdebe7"},
-    {file = "fonttools-4.55.4-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:5d8f925909256e62152e7c3e192655dbca3ab8c3cdef7d7b436732727e80feb6"},
-    {file = "fonttools-4.55.4-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:1a58af9b98e39bcd773aa352b4512be79b472830b799cb1d3cafb2b4796b71cd"},
-    {file = "fonttools-4.55.4-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:736d750d2ab4523067d8058e5294b40b01f2eee521e0fd401bec0d5e21e80b12"},
-    {file = "fonttools-4.55.4-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:1a9a2e7e8a9d3bfa9589db3e6c4e4c127fec252493924b2f87a67a25f9430057"},
-    {file = "fonttools-4.55.4-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:87824368e994af34a95cb4279a8c711e51974b3c28d052d39d768531cc9e8e59"},
-    {file = "fonttools-4.55.4-cp310-cp310-win32.whl", hash = "sha256:6c36dcbfe64bce38c4d4f1d436cdc6445e969eee96eb98d98be603b5abf8c3f2"},
-    {file = "fonttools-4.55.4-cp310-cp310-win_amd64.whl", hash = "sha256:3c53a467e5cf629acdbefc98b0f554859539fb6447bbeae4117b9ab51464ccc5"},
-    {file = "fonttools-4.55.4-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:1605b28165c785bf26c2cbd205dc0822463e3f9f56f187049eb214dc5f4a59cb"},
-    {file = "fonttools-4.55.4-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:d851d8b2fdb676507365d1430c3285d62c4039d0d7760d8cf2f2e5ea3aa19d73"},
-    {file = "fonttools-4.55.4-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:3fb3cf1cddf08cec0338f238f950cb76fabab23a324a579e3e1f9b2ef2578329"},
-    {file = "fonttools-4.55.4-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ddd3208b06186ca00fbd329c0d0fed5ba209c99017cc46e2c4ea42233c2fbd00"},
-    {file = "fonttools-4.55.4-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:9bd98819cb585a894dda9dcb337afeb2601abf17da17de7bfbfc1bc2e4a062c7"},
-    {file = "fonttools-4.55.4-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:4877376c10541e8dccf14876c8476d5082338fa5d21103894894382cc245144b"},
-    {file = "fonttools-4.55.4-cp311-cp311-win32.whl", hash = "sha256:3a5e466894ec6d8a009b0eb8e02a6eb26959a318d5b7a906280c26bdadce6423"},
-    {file = "fonttools-4.55.4-cp311-cp311-win_amd64.whl", hash = "sha256:f595129e6f9c6402965d6295fe8c18c1945d27af0f90bdb52ff426226e647afc"},
-    {file = "fonttools-4.55.4-cp312-cp312-macosx_10_13_universal2.whl", hash = "sha256:b3db72ad2d26a0e9ec694cbfb4485a8da9c095d29f66561cf935dbd19f3efcea"},
-    {file = "fonttools-4.55.4-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:87717808fd5953588c3ffaf512e8cab0e43c09c1da04e42ba87fa4c07d8170c7"},
-    {file = "fonttools-4.55.4-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:f49dac626ad5bc1a0147b88e6157e3211fd440d00007f0da6c9e5f91dd5cb88e"},
-    {file = "fonttools-4.55.4-cp312-cp312-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:a2d0ac8656ada8b604ae5da15d9aa075232f2181b95b51a3a2a55195222df7e7"},
-    {file = "fonttools-4.55.4-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:013c8b45873fa77a4ff6d25e43fecf1046cb7e8c6b32f1843117f98f3f8eac60"},
-    {file = "fonttools-4.55.4-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:94caad375d254a0332926512f06791f5e66c24a913ebecd6178b14f61d27c62f"},
-    {file = "fonttools-4.55.4-cp312-cp312-win32.whl", hash = "sha256:cb3eb4bf3a0c4e431e1ccab7a33ef4f1bb32657133fff4a61dc4fcbd54b94d29"},
-    {file = "fonttools-4.55.4-cp312-cp312-win_amd64.whl", hash = "sha256:6914269f6ff6b20c6b5a9b19d0b752880bd8ee218d9a7d6afe9960bbf1922d98"},
-    {file = "fonttools-4.55.4-cp313-cp313-macosx_10_13_universal2.whl", hash = "sha256:699dd32da7258a89939567a3d71b3f8decf84da54488a2526693f0d981a76479"},
-    {file = "fonttools-4.55.4-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:0f374b18ac04fbf78f20940418aee7882be3cdcb328ded80e16c3356499f64cf"},
-    {file = "fonttools-4.55.4-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:b18792529ca3c24259090b6faa60bd0bdfcc4a06312e8f06d6fccab007f07193"},
-    {file = "fonttools-4.55.4-cp313-cp313-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:8e91d25261ebc9ff2143b95e6272f46b9f28e260b8f40feda07c80b66ff7e61d"},
-    {file = "fonttools-4.55.4-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:2695781a897395d03504fd24b60c944726b5e7b7af9ea3d922f7319d70c6fc37"},
-    {file = "fonttools-4.55.4-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:21de3ef5b8e5361fd01d6aef2c09dda4ede139d6b3a1f5cf621d6bea48840dfd"},
-    {file = "fonttools-4.55.4-cp313-cp313-win32.whl", hash = "sha256:0ef33fda14e39aabb892a18ed16805b0b5b4e8a801fd1815a694be9dc7f30024"},
-    {file = "fonttools-4.55.4-cp313-cp313-win_amd64.whl", hash = "sha256:e953b1614e32b6da828ae7659c8f330a593b6c4b7a4a31f8f63c01b12f0d3680"},
-    {file = "fonttools-4.55.4-cp38-cp38-macosx_10_9_universal2.whl", hash = "sha256:e2d1bbcaf8ca8c60fbb029982197fbaa487559d5380f1c3098882c5ceb4311c7"},
-    {file = "fonttools-4.55.4-cp38-cp38-macosx_10_9_x86_64.whl", hash = "sha256:a885593dbcbfc250ff17831f7dc9316e95c3d046e6cd7ff7ab52ebf673bbf978"},
-    {file = "fonttools-4.55.4-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:02cd4ad9b3ab9f9c5b233b3bb6a96a036c9c0ef17487805b5e73cedf6439d188"},
-    {file = "fonttools-4.55.4-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:822d46676f794bb6cac055b43f5636792e2a360e18cf0f3a0333c21d79ec0f2d"},
-    {file = "fonttools-4.55.4-cp38-cp38-musllinux_1_2_aarch64.whl", hash = "sha256:7b195440fe14d8601053a51e06e13c94f725bf9f964611be99dc3cb65497ce8e"},
-    {file = "fonttools-4.55.4-cp38-cp38-musllinux_1_2_x86_64.whl", hash = "sha256:a0e0a0ec8cc4b8f82f9cf4efa26774dbd93433ba51b8f9bd2b214bf36c5638f6"},
-    {file = "fonttools-4.55.4-cp38-cp38-win32.whl", hash = "sha256:ca7e6047fbc995500e0b7459a04d5b92cafd7730b636d5f83334cd7eefdf95c7"},
-    {file = "fonttools-4.55.4-cp38-cp38-win_amd64.whl", hash = "sha256:0185983fcf49ae7a826cedc6f64d68b0434a5b7905d89e75bc95fced7fe118c1"},
-    {file = "fonttools-4.55.4-cp39-cp39-macosx_10_9_universal2.whl", hash = "sha256:dcc08dcb2be554073a72f3a8cecbc4226602ccdd0187b8f37a03a731cb931864"},
-    {file = "fonttools-4.55.4-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:7b9b414ce50f09cb692e97ff82b041ea1a21076ed9c1923206560c15ce9ad03a"},
-    {file = "fonttools-4.55.4-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:8807a1357d434ef1f4aed9bdfee7077f52dbc040b18ac98f6e417f69a48afbb5"},
-    {file = "fonttools-4.55.4-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:93a3ec7cba2e71edbc999ce3d48d34ef87cc30a36af6ff90dfc0dbc131f705fc"},
-    {file = "fonttools-4.55.4-cp39-cp39-musllinux_1_2_aarch64.whl", hash = "sha256:2964b9fe6b4a892a41a8a517bac232072a821cf2288fad1d19c6c1d19c34b0dd"},
-    {file = "fonttools-4.55.4-cp39-cp39-musllinux_1_2_x86_64.whl", hash = "sha256:0b9f4f032295adeb39a8c0eefb08a7b1e90f4b7571506e5d84bb923a7afa8247"},
-    {file = "fonttools-4.55.4-cp39-cp39-win32.whl", hash = "sha256:ee4e86280dc637a17e926cbdd32c2de148c013c3468777ae6e94c8b4449c8e93"},
-    {file = "fonttools-4.55.4-cp39-cp39-win_amd64.whl", hash = "sha256:82a03920f0f524abab375dcfac8926d9596986503ee00ae435bdd71b1498f214"},
-    {file = "fonttools-4.55.4-py3-none-any.whl", hash = "sha256:d07ad8f31038c6394a0945752458313367a0ef8125d284ee59f99e68393a3c2d"},
-    {file = "fonttools-4.55.4.tar.gz", hash = "sha256:9598af0af85073659facbe9612fcc56b071ef2f26e3819ebf9bd8c5d35f958c5"},
+    {file = "fonttools-4.55.8-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:d11600f5343092697d7434f3bf77a393c7ae74be206fe30e577b9a195fd53165"},
+    {file = "fonttools-4.55.8-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:c96f2506ce1a0beeaa9595f9a8b7446477eb133f40c0e41fc078744c28149f80"},
+    {file = "fonttools-4.55.8-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:9b5f05ef72e846e9f49ccdd74b9da4309901a4248434c63c1ee9321adcb51d65"},
+    {file = "fonttools-4.55.8-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ba45b637da80a262b55b7657aec68da2ac54b8ae7891cd977a5dbe5fd26db429"},
+    {file = "fonttools-4.55.8-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:edcffaeadba9a334c1c3866e275d7dd495465e7dbd296f688901bdbd71758113"},
+    {file = "fonttools-4.55.8-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:b9f9fce3c9b2196e162182ec5db8af8eb3acd0d76c2eafe9fdba5f370044e556"},
+    {file = "fonttools-4.55.8-cp310-cp310-win32.whl", hash = "sha256:f089e8da0990cfe2d67e81d9cf581ff372b48dc5acf2782701844211cd1f0eb3"},
+    {file = "fonttools-4.55.8-cp310-cp310-win_amd64.whl", hash = "sha256:01ea3901b0802fc5f9e854f5aeb5bc27770dd9dd24c28df8f74ba90f8b3f5915"},
+    {file = "fonttools-4.55.8-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:95f5a1d4432b3cea6571f5ce4f4e9b25bf36efbd61c32f4f90130a690925d6ee"},
+    {file = "fonttools-4.55.8-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:3d20f152de7625a0008ba1513f126daaaa0de3b4b9030aa72dd5c27294992260"},
+    {file = "fonttools-4.55.8-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:d5a3ff5bb95fd5a3962b2754f8435e6d930c84fc9e9921c51e802dddf40acd56"},
+    {file = "fonttools-4.55.8-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:b99d4fd2b6d0a00c7336c8363fccc7a11eccef4b17393af75ca6e77cf93ff413"},
+    {file = "fonttools-4.55.8-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:d637e4d33e46619c79d1a6c725f74d71b574cd15fb5bbb9b6f3eba8f28363573"},
+    {file = "fonttools-4.55.8-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:0f38bfb6b7a39c4162c3eb0820a0bdf8e3bdd125cd54e10ba242397d15e32439"},
+    {file = "fonttools-4.55.8-cp311-cp311-win32.whl", hash = "sha256:acfec948de41cd5e640d5c15d0200e8b8e7c5c6bb82afe1ca095cbc4af1188ee"},
+    {file = "fonttools-4.55.8-cp311-cp311-win_amd64.whl", hash = "sha256:604c805b41241b4880e2dc86cf2d4754c06777371c8299799ac88d836cb18c3b"},
+    {file = "fonttools-4.55.8-cp312-cp312-macosx_10_13_universal2.whl", hash = "sha256:63403ee0f2fa4e1de28e539f8c24f2bdca1d8ecb503fa9ea2d231d9f1e729809"},
+    {file = "fonttools-4.55.8-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:302e1003a760b222f711d5ba6d1ad7fd5f7f713eb872cd6a3eb44390bc9770af"},
+    {file = "fonttools-4.55.8-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:e72a7816ff8a759be9ca36ca46934f8ccf4383711ef597d9240306fe1878cb8d"},
+    {file = "fonttools-4.55.8-cp312-cp312-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:03c2b50b54e6e8b3564b232e57e8f58be217cf441cf0155745d9e44a76f9c30f"},
+    {file = "fonttools-4.55.8-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:a7230f7590f9570d26ee903b6a4540274494e200fae978df0d9325b7b9144529"},
+    {file = "fonttools-4.55.8-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:466a78984f0572305c3c48377f4e3f7f4e909f1209f45ef8e7041d5c8a744a56"},
+    {file = "fonttools-4.55.8-cp312-cp312-win32.whl", hash = "sha256:243cbfc0b7cb1c307af40e321f8343a48d0a080bc1f9466cf2b5468f776ef108"},
+    {file = "fonttools-4.55.8-cp312-cp312-win_amd64.whl", hash = "sha256:a19059aa892676822c1f05cb5a67296ecdfeb267fe7c47d4758f3e8e942c2b2a"},
+    {file = "fonttools-4.55.8-cp313-cp313-macosx_10_13_universal2.whl", hash = "sha256:332883b6280b9d90d2ba7e9e81be77cf2ace696161e60cdcf40cfcd2b3ed06fa"},
+    {file = "fonttools-4.55.8-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:6b8d7c149d47b47de7ec81763396c8266e5ebe2e0b14aa9c3ccf29e52260ab2f"},
+    {file = "fonttools-4.55.8-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:4dfae7c94987149bdaa0388e6c937566aa398fa0eec973b17952350a069cff4e"},
+    {file = "fonttools-4.55.8-cp313-cp313-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:a0fe12f06169af2fdc642d26a8df53e40adc3beedbd6ffedb19f1c5397b63afd"},
+    {file = "fonttools-4.55.8-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:f971aa5f50c22dc4b63a891503624ae2c77330429b34ead32f23c2260c5618cd"},
+    {file = "fonttools-4.55.8-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:708cb17b2590b7f6c6854999df0039ff1140dda9e6f56d67c3599ba6f968fab5"},
+    {file = "fonttools-4.55.8-cp313-cp313-win32.whl", hash = "sha256:cfe9cf30f391a0f2875247a3e5e44d8dcb61596e5cf89b360cdffec8a80e9961"},
+    {file = "fonttools-4.55.8-cp313-cp313-win_amd64.whl", hash = "sha256:1e10efc8ee10d6f1fe2931d41bccc90cd4b872f2ee4ff21f2231a2c293b2dbf8"},
+    {file = "fonttools-4.55.8-cp38-cp38-macosx_10_9_universal2.whl", hash = "sha256:9b6fcff4dc755b32faff955d989ee26394ddad3a90ea7d558db17a4633c8390c"},
+    {file = "fonttools-4.55.8-cp38-cp38-macosx_10_9_x86_64.whl", hash = "sha256:02c41322e5bdcb484b61b776fcea150215c83619b39c96aa0b44d4fd87bb5574"},
+    {file = "fonttools-4.55.8-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:9164f44add0acec0f12fce682824c040dc52e483bfe3838c37142897150c8364"},
+    {file = "fonttools-4.55.8-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:d2248ebfbcea0d0b3cb459d76a9f67f2eadc10ec0d07e9cadab8777d3f016bf2"},
+    {file = "fonttools-4.55.8-cp38-cp38-musllinux_1_2_aarch64.whl", hash = "sha256:3461347016c94cb42b36caa907e11565878c4c2c375604f3651d11dc06d1ab3e"},
+    {file = "fonttools-4.55.8-cp38-cp38-musllinux_1_2_x86_64.whl", hash = "sha256:67df1c3935838fb9e56f227d7f506c9043b149a4a3b667bef17929c7a1114d19"},
+    {file = "fonttools-4.55.8-cp38-cp38-win32.whl", hash = "sha256:cb121d6dd34625cece32234a5fa0359475bb118838b6b4295ffdb13b935edb04"},
+    {file = "fonttools-4.55.8-cp38-cp38-win_amd64.whl", hash = "sha256:285c1ac10c160fbdff6d05358230e66c4f98cbbf271f3ec7eb34e967771543e8"},
+    {file = "fonttools-4.55.8-cp39-cp39-macosx_10_9_universal2.whl", hash = "sha256:8abd135e427d88e461a4833c03cf96cfb9028c78c15d58123291f22398e25492"},
+    {file = "fonttools-4.55.8-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:65cb8f97eed7906dcf19bc2736b70c6239e9d7e77aad7c6110ba7239ae082e81"},
+    {file = "fonttools-4.55.8-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:450c354c04a6e12a3db968e915fe05730f79ff3d39560947ef8ee6eaa2ab2212"},
+    {file = "fonttools-4.55.8-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:2232012a1502b2b8ab4c6bc1d3524bfe90238c0c1a50ac94a0a2085aa87a58a5"},
+    {file = "fonttools-4.55.8-cp39-cp39-musllinux_1_2_aarch64.whl", hash = "sha256:d39f0c977639be0f9f5505d4c7c478236737f960c567a35f058649c056e41434"},
+    {file = "fonttools-4.55.8-cp39-cp39-musllinux_1_2_x86_64.whl", hash = "sha256:de78d6d0dbe32561ce059265437021f4746e56073c4799f0f1095828ae7232bd"},
+    {file = "fonttools-4.55.8-cp39-cp39-win32.whl", hash = "sha256:bf4b5b3496ddfdd4e57112e77ec51f1ab388d35ac17322c1248addb2eb0d429a"},
+    {file = "fonttools-4.55.8-cp39-cp39-win_amd64.whl", hash = "sha256:ccf8ae02918f431953d338db4d0a675a395faf82bab3a76025582cf32a2f3b7b"},
+    {file = "fonttools-4.55.8-py3-none-any.whl", hash = "sha256:07636dae94f7fe88561f9da7a46b13d8e3f529f87fdb221b11d85f91eabceeb7"},
+    {file = "fonttools-4.55.8.tar.gz", hash = "sha256:54d481d456dcd59af25d4a9c56b2c4c3f20e9620b261b84144e5950f33e8df17"},
 ]
 
 [package.dependencies]
@@ -1751,13 +1752,13 @@ test = ["flaky", "ipyparallel", "pre-commit", "pytest (>=7.0)", "pytest-asyncio
 
 [[package]]
 name = "ipython"
-version = "8.31.0"
+version = "8.32.0"
 description = "IPython: Productive Interactive Computing"
 optional = false
 python-versions = ">=3.10"
 files = [
-    {file = "ipython-8.31.0-py3-none-any.whl", hash = "sha256:46ec58f8d3d076a61d128fe517a51eb730e3aaf0c184ea8c17d16e366660c6a6"},
-    {file = "ipython-8.31.0.tar.gz", hash = "sha256:b6a2274606bec6166405ff05e54932ed6e5cfecaca1fc05f2cacde7bb074d70b"},
+    {file = "ipython-8.32.0-py3-none-any.whl", hash = "sha256:cae85b0c61eff1fc48b0a8002de5958b6528fa9c8defb1894da63f42613708aa"},
+    {file = "ipython-8.32.0.tar.gz", hash = "sha256:be2c91895b0b9ea7ba49d33b23e2040c352b33eb6a519cca7ce6e0c743444251"},
 ]
 
 [package.dependencies]
@@ -2024,17 +2025,18 @@ test = ["ipykernel", "pre-commit", "pytest (<8)", "pytest-cov", "pytest-timeout"
 
 [[package]]
 name = "jupyter-events"
-version = "0.11.0"
+version = "0.12.0"
 description = "Jupyter Event System library"
 optional = false
 python-versions = ">=3.9"
 files = [
-    {file = "jupyter_events-0.11.0-py3-none-any.whl", hash = "sha256:36399b41ce1ca45fe8b8271067d6a140ffa54cec4028e95491c93b78a855cacf"},
-    {file = "jupyter_events-0.11.0.tar.gz", hash = "sha256:c0bc56a37aac29c1fbc3bcfbddb8c8c49533f9cf11f1c4e6adadba936574ab90"},
+    {file = "jupyter_events-0.12.0-py3-none-any.whl", hash = "sha256:6464b2fa5ad10451c3d35fabc75eab39556ae1e2853ad0c0cc31b656731a97fb"},
+    {file = "jupyter_events-0.12.0.tar.gz", hash = "sha256:fc3fce98865f6784c9cd0a56a20644fc6098f21c8c33834a8d9fe383c17e554b"},
 ]
 
 [package.dependencies]
 jsonschema = {version = ">=4.18.0", extras = ["format-nongpl"]}
+packaging = "*"
 python-json-logger = ">=2.0.4"
 pyyaml = ">=5.3"
 referencing = "*"
@@ -2677,13 +2679,13 @@ dill = ">=0.3.8"
 
 [[package]]
 name = "narwhals"
-version = "1.24.1"
+version = "1.25.1"
 description = "Extremely lightweight compatibility layer between dataframe libraries"
 optional = false
 python-versions = ">=3.8"
 files = [
-    {file = "narwhals-1.24.1-py3-none-any.whl", hash = "sha256:d8983fe14851c95d60576ddca37c094bd4ed24ab9ea98396844fb20ad9aaf184"},
-    {file = "narwhals-1.24.1.tar.gz", hash = "sha256:b09b8253d945f23cdb683a84685abf3afb9f96114d89e9f35dc876e143f65007"},
+    {file = "narwhals-1.25.1-py3-none-any.whl", hash = "sha256:a1838f2725523da54c093849e93a8b2a57d2310f0bbc26be35d223f5eef60417"},
+    {file = "narwhals-1.25.1.tar.gz", hash = "sha256:9c0e27be46e186526878286b442a3dd2ee9fe723456457feff42316288732b96"},
 ]
 
 [package.extras]
@@ -4113,17 +4115,18 @@ files = [
 
 [[package]]
 name = "pywinpty"
-version = "2.0.14"
+version = "2.0.15"
 description = "Pseudo terminal support for Windows from Python."
 optional = false
-python-versions = ">=3.8"
+python-versions = ">=3.9"
 files = [
-    {file = "pywinpty-2.0.14-cp310-none-win_amd64.whl", hash = "sha256:0b149c2918c7974f575ba79f5a4aad58bd859a52fa9eb1296cc22aa412aa411f"},
-    {file = "pywinpty-2.0.14-cp311-none-win_amd64.whl", hash = "sha256:cf2a43ac7065b3e0dc8510f8c1f13a75fb8fde805efa3b8cff7599a1ef497bc7"},
-    {file = "pywinpty-2.0.14-cp312-none-win_amd64.whl", hash = "sha256:55dad362ef3e9408ade68fd173e4f9032b3ce08f68cfe7eacb2c263ea1179737"},
-    {file = "pywinpty-2.0.14-cp313-none-win_amd64.whl", hash = "sha256:074fb988a56ec79ca90ed03a896d40707131897cefb8f76f926e3834227f2819"},
-    {file = "pywinpty-2.0.14-cp39-none-win_amd64.whl", hash = "sha256:5725fd56f73c0531ec218663bd8c8ff5acc43c78962fab28564871b5fce053fd"},
-    {file = "pywinpty-2.0.14.tar.gz", hash = "sha256:18bd9529e4a5daf2d9719aa17788ba6013e594ae94c5a0c27e83df3278b0660e"},
+    {file = "pywinpty-2.0.15-cp310-cp310-win_amd64.whl", hash = "sha256:8e7f5de756a615a38b96cd86fa3cd65f901ce54ce147a3179c45907fa11b4c4e"},
+    {file = "pywinpty-2.0.15-cp311-cp311-win_amd64.whl", hash = "sha256:9a6bcec2df2707aaa9d08b86071970ee32c5026e10bcc3cc5f6f391d85baf7ca"},
+    {file = "pywinpty-2.0.15-cp312-cp312-win_amd64.whl", hash = "sha256:83a8f20b430bbc5d8957249f875341a60219a4e971580f2ba694fbfb54a45ebc"},
+    {file = "pywinpty-2.0.15-cp313-cp313-win_amd64.whl", hash = "sha256:ab5920877dd632c124b4ed17bc6dd6ef3b9f86cd492b963ffdb1a67b85b0f408"},
+    {file = "pywinpty-2.0.15-cp313-cp313t-win_amd64.whl", hash = "sha256:a4560ad8c01e537708d2790dbe7da7d986791de805d89dd0d3697ca59e9e4901"},
+    {file = "pywinpty-2.0.15-cp39-cp39-win_amd64.whl", hash = "sha256:d261cd88fcd358cfb48a7ca0700db3e1c088c9c10403c9ebc0d8a8b57aa6a117"},
+    {file = "pywinpty-2.0.15.tar.gz", hash = "sha256:312cf39153a8736c617d45ce8b6ad6cd2107de121df91c455b10ce6bba7a39b2"},
 ]
 
 [[package]]
@@ -4985,13 +4988,13 @@ full = ["httpx (>=0.27.0,<0.29.0)", "itsdangerous", "jinja2", "python-multipart
 
 [[package]]
 name = "streamlit"
-version = "1.41.1"
+version = "1.42.0"
 description = "A faster way to build and share data apps"
 optional = false
 python-versions = "!=3.9.7,>=3.9"
 files = [
-    {file = "streamlit-1.41.1-py2.py3-none-any.whl", hash = "sha256:0def00822480071d642e6df36cd63c089f991da3a69fd9eb4ab8f65ce27de4e0"},
-    {file = "streamlit-1.41.1.tar.gz", hash = "sha256:6626d32b098ba1458b71eebdd634c62af2dd876380e59c4b6a1e828a39d62d69"},
+    {file = "streamlit-1.42.0-py2.py3-none-any.whl", hash = "sha256:edf333fd3525b7c64b19e1156b483a1a93cbdb09a3a06f26478388d68f971090"},
+    {file = "streamlit-1.42.0.tar.gz", hash = "sha256:8c48494ccfad33e7d0bc5873151800b203cb71203bfd42bc7418940710ca4970"},
 ]
 
 [package.dependencies]
@@ -5012,11 +5015,11 @@ rich = ">=10.14.0,<14"
 tenacity = ">=8.1.0,<10"
 toml = ">=0.10.1,<2"
 tornado = ">=6.0.3,<7"
-typing-extensions = ">=4.3.0,<5"
+typing-extensions = ">=4.4.0,<5"
 watchdog = {version = ">=2.1.5,<7", markers = "platform_system != \"Darwin\""}
 
 [package.extras]
-snowflake = ["snowflake-connector-python (>=2.8.0)", "snowflake-snowpark-python[modin] (>=1.17.0)"]
+snowflake = ["snowflake-connector-python (>=3.3.0)", "snowflake-snowpark-python[modin] (>=1.17.0)"]
 
 [[package]]
 name = "surya-ocr"
@@ -5683,13 +5686,13 @@ files = [
 
 [[package]]
 name = "xlsxwriter"
-version = "3.2.1"
+version = "3.2.2"
 description = "A Python module for creating Excel XLSX files."
 optional = false
 python-versions = ">=3.6"
 files = [
-    {file = "XlsxWriter-3.2.1-py3-none-any.whl", hash = "sha256:7e8f7c60b7a1660ef791d46ab5de78469cb978b991ca841af61f5832d2f9f4fe"},
-    {file = "XlsxWriter-3.2.1.tar.gz", hash = "sha256:97618759cb264fb6a93397f660cca156ffa9561743b1823dafb60dc4474e1902"},
+    {file = "XlsxWriter-3.2.2-py3-none-any.whl", hash = "sha256:272ce861e7fa5e82a4a6ebc24511f2cb952fde3461f6c6e1a1e81d3272db1471"},
+    {file = "xlsxwriter-3.2.2.tar.gz", hash = "sha256:befc7f92578a85fed261639fb6cde1fd51b79c5e854040847dde59d4317077dc"},
 ]
 
 [[package]]
@@ -6006,4 +6009,4 @@ test = ["pytest"]
 [metadata]
 lock-version = "2.0"
 python-versions = "^3.10"
-content-hash = "9730ffc5216b8e9eb6d4c59573f4d382a160480f1a3f7fcae290d52d4e6f8a28"
+content-hash = "78001740906e61993933c0f6708f0326fd5206e66f4a27fc48c1c45b38f82c1d"

From 4d105da58823113a98f5ab13538cad48a04dd32a Mon Sep 17 00:00:00 2001
From: Moses Paul R <iammosespaulr@gmail.com>
Date: Thu, 6 Feb 2025 09:15:41 +0000
Subject: [PATCH 07/46] add epub and html support

---
 marker/providers/epub.py     | 115 +++++++++++++++++++++++++++++++++++
 marker/providers/html.py     |  37 +++++++++++
 marker/providers/registry.py |  13 ++++
 poetry.lock                  |  16 ++++-
 pyproject.toml               |   1 +
 5 files changed, 181 insertions(+), 1 deletion(-)
 create mode 100644 marker/providers/epub.py
 create mode 100644 marker/providers/html.py

diff --git a/marker/providers/epub.py b/marker/providers/epub.py
new file mode 100644
index 00000000..f50e0ae9
--- /dev/null
+++ b/marker/providers/epub.py
@@ -0,0 +1,115 @@
+import base64
+import logging
+import os
+
+import ebooklib
+from bs4 import BeautifulSoup
+from ebooklib import epub
+from weasyprint import CSS, HTML
+
+from marker.providers.pdf import PdfProvider
+
+logging.getLogger('fontTools.subset').setLevel(logging.ERROR)
+logging.getLogger('fontTools.ttLib.ttFont').setLevel(logging.ERROR)
+logging.getLogger('weasyprint').setLevel(logging.ERROR)
+
+css = '''
+@page {
+    size: A4;
+    margin: 2cm;
+}
+
+img {
+    max-width: 100%;
+    max-height: 25cm;
+    object-fit: contain;
+    margin: 12pt auto;
+}
+
+div, p {
+    max-width: 100%;
+    word-break: break-word;
+    font-size: 10pt;
+}
+
+table {
+    width: 100%;
+    border-collapse: collapse;
+    break-inside: auto;
+    font-size: 10pt;
+}
+
+tr {
+    break-inside: avoid;
+    page-break-inside: avoid;
+}
+
+td {
+    border: 0.75pt solid #000;
+    padding: 6pt;
+}
+'''
+
+
+class EpubProvider(PdfProvider):
+    def __init__(self, filepath: str, config=None):
+        home_dir = os.path.expanduser("~")
+        rel_path = os.path.relpath(filepath, home_dir)
+        base_name, _ = os.path.splitext(rel_path)
+        self.temp_pdf_path = os.path.join('/tmp', f"{base_name}.pdf")
+
+        # Convert Epub to PDF
+        try:
+            self.convert_epub_to_pdf(filepath)
+        except Exception as e:
+            raise RuntimeError(f"Failed to convert {filepath} to PDF: {e}")
+
+        # Initialize the PDF provider with the temp pdf path
+        super().__init__(self.temp_pdf_path, config)
+
+    def __del__(self):
+        if os.path.exists(self.temp_pdf_path):
+            print(f"Deleting temporary PDF file: {self.temp_pdf_path}")
+            os.remove(self.temp_pdf_path)
+
+    def convert_epub_to_pdf(self, filepath):
+        ebook = epub.read_epub(filepath)
+
+        styles = []
+        html_content = ""
+        img_tags = {}
+
+        for item in ebook.get_items():
+            if item.get_type() == ebooklib.ITEM_IMAGE:
+                img_data = base64.b64encode(item.get_content()).decode("utf-8")
+                img_tags[item.file_name] = f'data:{item.media_type};base64,{img_data}'
+            elif item.get_type() == ebooklib.ITEM_STYLE:
+                styles.append(item.get_content().decode('utf-8'))
+
+        for item in ebook.get_items():
+            if item.get_type() == ebooklib.ITEM_DOCUMENT:
+                html_content += item.get_content().decode("utf-8")
+
+        soup = BeautifulSoup(html_content, 'html.parser')
+        for img in soup.find_all('img'):
+            src = img.get('src')
+            if src:
+                normalized_src = src.replace('../', '')
+                if normalized_src in img_tags:
+                    img['src'] = img_tags[normalized_src]
+
+        for image in soup.find_all('image'):
+            src = image.get('xlink:href')
+            if src:
+                normalized_src = src.replace('../', '')
+                if normalized_src in img_tags:
+                    image['xlink:href'] = img_tags[normalized_src]
+
+        html_content = str(soup)
+        full_style = ''.join([css])# + styles)
+
+        # we convert the epub to HTML
+        result = HTML(string=html_content, base_url=filepath).write_pdf(
+            self.temp_pdf_path,
+            stylesheets=[CSS(string=full_style)]
+        )
diff --git a/marker/providers/html.py b/marker/providers/html.py
new file mode 100644
index 00000000..47a1a9ed
--- /dev/null
+++ b/marker/providers/html.py
@@ -0,0 +1,37 @@
+import logging
+import os
+
+from weasyprint import HTML
+
+from marker.providers.pdf import PdfProvider
+
+logging.getLogger('fontTools.subset').setLevel(logging.ERROR)
+logging.getLogger('fontTools.ttLib.ttFont').setLevel(logging.ERROR)
+
+class HTMLProvider(PdfProvider):
+    def __init__(self, filepath: str, config=None):
+        home_dir = os.path.expanduser("~")
+        rel_path = os.path.relpath(filepath, home_dir)
+        base_name, _ = os.path.splitext(rel_path)
+        self.temp_pdf_path = os.path.join('/tmp', f"{base_name}.pdf")
+
+        # Convert HTML to PDF
+        try:
+            self.convert_html_to_pdf(filepath)
+        except Exception as e:
+            raise RuntimeError(f"Failed to convert {filepath} to PDF: {e}")
+
+        # Initialize the PDF provider with the temp pdf path
+        super().__init__(self.temp_pdf_path, config)
+
+    def __del__(self):
+        if os.path.exists(self.temp_pdf_path):
+            print(f"Deleting temporary PDF file: {self.temp_pdf_path}")
+            os.remove(self.temp_pdf_path)
+
+    def convert_html_to_pdf(self, filepath: str):
+        with open(filepath, "rb") as html_file:
+            # we convert the html to PDF
+            HTML(string=html_file.read()).write_pdf(
+                self.temp_pdf_path,
+            )
diff --git a/marker/providers/registry.py b/marker/providers/registry.py
index f3dfb812..8cabc588 100644
--- a/marker/providers/registry.py
+++ b/marker/providers/registry.py
@@ -1,8 +1,11 @@
 import filetype
 import filetype.match as match
+from bs4 import BeautifulSoup
 from filetype.types import archive, document
 
 from marker.providers.document import DocumentProvider
+from marker.providers.epub import EpubProvider
+from marker.providers.html import HTMLProvider
 from marker.providers.image import ImageProvider
 from marker.providers.pdf import PdfProvider
 from marker.providers.powerpoint import PowerPointProvider
@@ -14,6 +17,8 @@ def provider_from_filepath(filepath: str):
         return ImageProvider
     if match(filepath, (archive.Pdf(),)) is not None:
         return PdfProvider
+    if match(filepath, (archive.Epub(),)) is not None:
+        return EpubProvider
     if match(
             filepath, (
                 document.Doc(),
@@ -36,4 +41,12 @@ def provider_from_filepath(filepath: str):
             )) is not None:
         return PowerPointProvider
 
+    try:
+        soup = BeautifulSoup(open(filepath, 'r').read(), 'html.parser')
+        # Check if there are any HTML tags
+        if bool(soup.find()):
+            return HTMLProvider
+    except:
+        pass
+
     return PdfProvider
diff --git a/poetry.lock b/poetry.lock
index ca581c10..6df1ba98 100644
--- a/poetry.lock
+++ b/poetry.lock
@@ -975,6 +975,20 @@ files = [
     {file = "Distance-0.1.3.tar.gz", hash = "sha256:60807584f5b6003f5c521aa73f39f51f631de3be5cccc5a1d67166fcbf0d4551"},
 ]
 
+[[package]]
+name = "ebooklib"
+version = "0.18"
+description = "Ebook library which can handle EPUB2/EPUB3 and Kindle format"
+optional = false
+python-versions = "*"
+files = [
+    {file = "EbookLib-0.18.tar.gz", hash = "sha256:38562643a7bc94d9bf56e9930b4927e4e93b5d1d0917f697a6454db5a1c1a533"},
+]
+
+[package.dependencies]
+lxml = "*"
+six = "*"
+
 [[package]]
 name = "et-xmlfile"
 version = "2.0.0"
@@ -6009,4 +6023,4 @@ test = ["pytest"]
 [metadata]
 lock-version = "2.0"
 python-versions = "^3.10"
-content-hash = "78001740906e61993933c0f6708f0326fd5206e66f4a27fc48c1c45b38f82c1d"
+content-hash = "1cba9dec2bd9279e646e67b6faeea8fb55c08d9860d7dffc0a44c5f6b7e226d8"
diff --git a/pyproject.toml b/pyproject.toml
index c1a11c74..e353aed3 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -40,6 +40,7 @@ mammoth = "^1.9.0"
 weasyprint = "^63.1"
 openpyxl = "^3.1.5"
 python-pptx = "^1.0.2"
+ebooklib = "^0.18"
 
 [tool.poetry.group.dev.dependencies]
 jupyter = "^1.0.0"

From 2f71b200ce327c19bb515dc89a692537758cb662 Mon Sep 17 00:00:00 2001
From: Moses Paul R <iammosespaulr@gmail.com>
Date: Thu, 6 Feb 2025 09:32:39 +0000
Subject: [PATCH 08/46] simplify tempfiles

---
 marker/providers/document.py    |  8 ++++----
 marker/providers/epub.py        | 10 +++++-----
 marker/providers/html.py        |  9 +++++----
 marker/providers/powerpoint.py  |  8 ++++----
 marker/providers/spreadsheet.py | 10 +++++-----
 5 files changed, 23 insertions(+), 22 deletions(-)

diff --git a/marker/providers/document.py b/marker/providers/document.py
index 025800d6..e68a7363 100644
--- a/marker/providers/document.py
+++ b/marker/providers/document.py
@@ -2,6 +2,7 @@
 import logging
 import os
 import re
+import tempfile
 from io import BytesIO
 
 import mammoth
@@ -53,10 +54,9 @@
 
 class DocumentProvider(PdfProvider):
     def __init__(self, filepath: str, config=None):
-        home_dir = os.path.expanduser("~")
-        rel_path = os.path.relpath(filepath, home_dir)
-        base_name, _ = os.path.splitext(rel_path)
-        self.temp_pdf_path = os.path.join('/tmp', f"{base_name}.pdf")
+        temp_pdf = tempfile.NamedTemporaryFile(delete=False, suffix=f".pdf")
+        self.temp_pdf_path = temp_pdf.name
+        temp_pdf.close()
 
         # Convert DOCX to PDF
         try:
diff --git a/marker/providers/epub.py b/marker/providers/epub.py
index f50e0ae9..74627a8a 100644
--- a/marker/providers/epub.py
+++ b/marker/providers/epub.py
@@ -1,6 +1,7 @@
 import base64
 import logging
 import os
+import tempfile
 
 import ebooklib
 from bs4 import BeautifulSoup
@@ -53,10 +54,9 @@
 
 class EpubProvider(PdfProvider):
     def __init__(self, filepath: str, config=None):
-        home_dir = os.path.expanduser("~")
-        rel_path = os.path.relpath(filepath, home_dir)
-        base_name, _ = os.path.splitext(rel_path)
-        self.temp_pdf_path = os.path.join('/tmp', f"{base_name}.pdf")
+        temp_pdf = tempfile.NamedTemporaryFile(delete=False, suffix=f".pdf")
+        self.temp_pdf_path = temp_pdf.name
+        temp_pdf.close()
 
         # Convert Epub to PDF
         try:
@@ -106,7 +106,7 @@ def convert_epub_to_pdf(self, filepath):
                     image['xlink:href'] = img_tags[normalized_src]
 
         html_content = str(soup)
-        full_style = ''.join([css])# + styles)
+        full_style = ''.join([css])  # + styles)
 
         # we convert the epub to HTML
         result = HTML(string=html_content, base_url=filepath).write_pdf(
diff --git a/marker/providers/html.py b/marker/providers/html.py
index 47a1a9ed..eab4c908 100644
--- a/marker/providers/html.py
+++ b/marker/providers/html.py
@@ -1,5 +1,6 @@
 import logging
 import os
+import tempfile
 
 from weasyprint import HTML
 
@@ -8,12 +9,12 @@
 logging.getLogger('fontTools.subset').setLevel(logging.ERROR)
 logging.getLogger('fontTools.ttLib.ttFont').setLevel(logging.ERROR)
 
+
 class HTMLProvider(PdfProvider):
     def __init__(self, filepath: str, config=None):
-        home_dir = os.path.expanduser("~")
-        rel_path = os.path.relpath(filepath, home_dir)
-        base_name, _ = os.path.splitext(rel_path)
-        self.temp_pdf_path = os.path.join('/tmp', f"{base_name}.pdf")
+        temp_pdf = tempfile.NamedTemporaryFile(delete=False, suffix=f".pdf")
+        self.temp_pdf_path = temp_pdf.name
+        temp_pdf.close()
 
         # Convert HTML to PDF
         try:
diff --git a/marker/providers/powerpoint.py b/marker/providers/powerpoint.py
index 433f7f4b..2a1febd1 100644
--- a/marker/providers/powerpoint.py
+++ b/marker/providers/powerpoint.py
@@ -1,5 +1,6 @@
 import base64
 import os
+import tempfile
 import traceback
 
 from pptx import Presentation
@@ -43,10 +44,9 @@ class PowerPointProvider(PdfProvider):
     include_slide_number: bool = False
 
     def __init__(self, filepath: str, config=None):
-        home_dir = os.path.expanduser("~")
-        rel_path = os.path.relpath(filepath, home_dir)
-        base_name, _ = os.path.splitext(rel_path)
-        self.temp_pdf_path = os.path.join('/tmp', f"{base_name}.pdf")
+        temp_pdf = tempfile.NamedTemporaryFile(delete=False, suffix=f".pdf")
+        self.temp_pdf_path = temp_pdf.name
+        temp_pdf.close()
 
         # Convert PPTX to PDF
         try:
diff --git a/marker/providers/spreadsheet.py b/marker/providers/spreadsheet.py
index 50ae30e2..01fe5c61 100644
--- a/marker/providers/spreadsheet.py
+++ b/marker/providers/spreadsheet.py
@@ -1,10 +1,11 @@
 import os
+import tempfile
 
 from openpyxl import load_workbook
 from openpyxl.worksheet.worksheet import Worksheet
+from weasyprint import CSS, HTML
 
 from marker.providers.pdf import PdfProvider
-from weasyprint import CSS, HTML
 
 css = '''
 @page {
@@ -33,10 +34,9 @@
 
 class SpreadSheetProvider(PdfProvider):
     def __init__(self, filepath: str, config=None):
-        home_dir = os.path.expanduser("~")
-        rel_path = os.path.relpath(filepath, home_dir)
-        base_name, _ = os.path.splitext(rel_path)
-        self.temp_pdf_path = os.path.join('/tmp', f"{base_name}.pdf")
+        temp_pdf = tempfile.NamedTemporaryFile(delete=False, suffix=f".pdf")
+        self.temp_pdf_path = temp_pdf.name
+        temp_pdf.close()
 
         # Convert XLSX to PDF
         try:

From d142598edf4d7901d188ca7a6858a98137b6405a Mon Sep 17 00:00:00 2001
From: Vik Paruchuri <vik.paruchuri@gmail.com>
Date: Thu, 20 Feb 2025 16:32:19 -0500
Subject: [PATCH 09/46] Iterate on inline math

---
 marker/builders/line.py                 |  48 +++----
 marker/converters/pdf.py                |   3 +
 marker/processors/llm/llm_inlinemath.py | 166 ++++++++++++++++++++++++
 marker/processors/llm/llm_text.py       |   3 +-
 marker/renderers/__init__.py            |  11 ++
 marker/renderers/html.py                |   1 +
 6 files changed, 198 insertions(+), 34 deletions(-)
 create mode 100644 marker/processors/llm/llm_inlinemath.py

diff --git a/marker/builders/line.py b/marker/builders/line.py
index 579f8480..ef0da08a 100644
--- a/marker/builders/line.py
+++ b/marker/builders/line.py
@@ -1,4 +1,6 @@
+from collections import defaultdict
 from copy import deepcopy
+from itertools import chain
 from typing import Annotated, List, Optional, Tuple
 
 import numpy as np
@@ -71,14 +73,14 @@ class LineBuilder(BaseBuilder):
         float,
         "The minimum overlap of a line with an inline math box to consider as a match"
     ] = 0.
+    line_inline_min_overlap_pct: Annotated[
+        float,
+        "The percentage of a provider line that has to be covered by a math line."
+    ] = .3
     line_text_overlap_threshold: Annotated[
         float,
         "The minimum overlap of an equation with a text line to consider as a match"
     ] = .5
-    inline_math_minimum_area: Annotated[
-        float,
-        "The minimum area for an inline math block, in pixels."
-    ] = 20
     inline_math_line_vertical_merge_threshold: Annotated[
         int,
         "The maximum pixel distance between y1s for two lines to be merged"
@@ -365,7 +367,7 @@ def determine_math_lines(
             max_overlap = np.max(overlap_row) / inline_box.area
 
             # Avoid small or nonoverlapping inline math regions
-            if max_overlap <= self.line_inline_math_overlap_threshold or inline_box.area < self.inline_math_minimum_area:
+            if max_overlap <= self.line_inline_math_overlap_threshold:
                 continue
 
             # Ignore vertical lines
@@ -401,45 +403,25 @@ def merge_provider_lines_inline_math(
         provider_line_boxes = [p.line.polygon.bbox for _, p in horizontal_provider_lines]
         math_line_boxes = [PolygonBox(polygon=m.polygon).rescale(image_size, page_size).bbox for m in inline_math_lines]
 
-        overlaps = matrix_intersection_area(math_line_boxes, provider_line_boxes)
+        overlaps = matrix_intersection_area(provider_line_boxes, math_line_boxes)
 
         # Find potential merges
-        merge_lines = []
-        for i in range(len(math_line_boxes)):
-            merge_line = []
-            math_line_polygon = PolygonBox(polygon=inline_math_lines[i].polygon).rescale(image_size, page_size)
-            max_overlap = np.max(overlaps[i])
-            if max_overlap <= self.line_inline_math_overlap_threshold:
+        merge_lines = defaultdict(list)
+        for i in range(len(provider_line_boxes)):
+            max_overlap_pct = np.max(overlaps[i]) / horizontal_provider_lines[i][1].line.polygon.area
+            if max_overlap_pct <= self.line_inline_min_overlap_pct:
                 continue
 
             best_overlap = np.argmax(overlaps[i])
             best_overlap_line = horizontal_provider_lines[best_overlap]
-            best_overlap_y1 = best_overlap_line[1].line.polygon.y_start
-
-            nonzero_idxs = np.nonzero(overlaps[i] > self.line_inline_math_overlap_threshold)[0]
-            for idx in nonzero_idxs:
-                provider_idx, provider_line = horizontal_provider_lines[idx]
-                provider_line_y1 = provider_line.line.polygon.y_start
-
-                should_merge_line = False
-                if abs(provider_line_y1 - best_overlap_y1) <= self.inline_math_line_vertical_merge_threshold:
-                    should_merge_line = True
-
-                line_overlaps = self.find_overlapping_math_chars(provider_line, math_line_polygon, remove_chars=not should_merge_line)
-
-                # Do not merge if too far above/below (but remove characters)
-                if line_overlaps and should_merge_line:
-                    # Add the index of the provider line to the merge line
-                    merge_line.append(provider_idx)
 
-            if len(merge_line) > 0:
-                merge_lines.append(merge_line)
+            merge_lines[best_overlap].append(i)
 
         # Handle the merging
         already_merged = set()
-        potential_merges = set([m for merge_line in merge_lines for m in merge_line])
+        potential_merges = set(chain.from_iterable(merge_lines.values()))
         out_provider_lines = [(i, p) for i, p in enumerate(provider_lines) if i not in potential_merges]
-        for merge_section in merge_lines:
+        for merge_section in merge_lines.values():
             merge_section = [m for m in merge_section if m not in already_merged]
             if len(merge_section) == 0:
                 continue
diff --git a/marker/converters/pdf.py b/marker/converters/pdf.py
index 400714ee..c324ca2d 100644
--- a/marker/converters/pdf.py
+++ b/marker/converters/pdf.py
@@ -1,5 +1,7 @@
 import os
 
+from marker.processors.llm.llm_inlinemath import LLMInlineMathProcessor
+
 os.environ["TOKENIZERS_PARALLELISM"] = "false"  # disables a tokenizers warning
 
 from collections import defaultdict
@@ -82,6 +84,7 @@ class PdfConverter(BaseConverter):
         LLMImageDescriptionProcessor,
         LLMEquationProcessor,
         LLMHandwritingProcessor,
+        LLMInlineMathProcessor,
         ReferenceProcessor,
         DebugProcessor,
     )
diff --git a/marker/processors/llm/llm_inlinemath.py b/marker/processors/llm/llm_inlinemath.py
new file mode 100644
index 00000000..426e2a92
--- /dev/null
+++ b/marker/processors/llm/llm_inlinemath.py
@@ -0,0 +1,166 @@
+import json
+from concurrent.futures import ThreadPoolExecutor, as_completed
+from typing import List, Tuple, Annotated
+
+from pydantic import BaseModel
+from tqdm import tqdm
+
+from marker.processors.llm import BaseLLMComplexBlockProcessor
+
+from marker.processors.util import text_to_spans
+from marker.schema import BlockTypes
+from marker.schema.blocks import Block
+from marker.schema.document import Document
+from marker.schema.groups import PageGroup
+from marker.schema.registry import get_block_class
+
+
+class LLMInlineMathProcessor(BaseLLMComplexBlockProcessor):
+    redo_inline_math: Annotated[
+        bool,
+        "If True, the inline math will be re-done, otherwise it will be left as is."
+    ] = False
+
+    block_types = (BlockTypes.TextInlineMath,)
+    text_math_rewriting_prompt = """You are a text correction expert specializing in accurately reproducing text from images.
+You will receive an image of a text block and a set of extracted lines corresponding to the text in the image.
+Your task is to correct any errors in the extracted lines, including math, formatting, and other inaccuracies, and output the corrected lines in a JSON format.
+The number of output lines MUST match the number of input lines.  There are {input_line_count} input lines. Stay as faithful to the original text as possible.
+
+**Instructions:**
+
+1. Carefully examine the provided text block image .
+2. Analyze the extracted lines.
+3. For each extracted line, compare it to the corresponding line in the image.
+4. Correct any errors in the extracted line, including:
+    * Inline math: Ensure all mathematical expressions are correctly formatted and rendered.  Surround them with <math>...</math> tags.
+    * Formatting: Maintain consistent formatting with the text block image, including spacing, indentation, and special characters.
+    * Other inaccuracies:  If the image is handwritten then you may correct any spelling errors, or other discrepancies.
+5. Do not remove any formatting i.e bold, italics, math, superscripts, subscripts, etc from the extracted lines unless it is necessary to correct an error.
+6. Ensure that inline math is properly with inline math tags, like <math>...</math>.
+7. The number of corrected lines in the output MUST equal the number of extracted lines provided in the input. Do not add or remove lines.
+8. Output the corrected lines in JSON format, as shown in the example below.  Each line should be in HTML format. Only use the math, br, a, i, b, sup, sub, and span tags.
+9. You absolutely cannot remove any <a href='#...'>...</a> tags, those are extremely important for references and are coming directly from the document, you MUST always preserve them.
+
+**Example:**
+
+Input:
+```
+{
+ "extracted_lines": [
+  "Adversarial training (AT) <a href='#page-9-1'>[23]</a>, which aims to minimize\n",
+  "the model's risk under the worst-case perturbations, is cur-\n",
+  "rently the most effective approach for improving the robust-\n",
+  "ness of deep neural networks. For a given neural network\n",
+  "f(x, w) with parameters w, the optimization objective of\n",
+  "AT can be formulated as follows:\n"
+ ]
+}
+```
+
+Output:
+
+```json
+{
+ "corrected_lines": [
+  "Adversarial training (AT) <a href='#page-9-1'>[23]</a>, which aims to minimize\n",
+  "the model's risk under the worst-case perturbations, is cur-\n",
+  "rently the most effective approach for improving the robust-\n",
+  "ness of deep neural networks. For a given neural network\n",
+  "<math>f(x, w)</math> with parameters <math>w</math>, the optimization objective of\n",
+  "AT can be formulated as follows:\n"
+ ]
+}
+```
+
+**Input:**
+```json
+{extracted_lines}
+```
+"""
+
+    def rewrite_blocks(self, document: Document):
+        if not self.redo_inline_math:
+            return
+
+        inline_blocks = [
+            (page, block)
+            for page in document.pages
+            for block in page.contained_blocks(document, self.block_types)
+        ]
+        detected_blocks = [
+            (page, block)
+            for page in document.pages
+            for block in page.contained_blocks(document, (BlockTypes.Text, BlockTypes.Caption, BlockTypes.SectionHeader))
+            if any([b.formats and "math" in b.formats for b in block.contained_blocks(document, (BlockTypes.Line,))])
+        ]
+        inference_blocks = inline_blocks + detected_blocks
+        # Don't show progress if there are no blocks to process
+        total_blocks = len(inference_blocks)
+        if total_blocks == 0:
+            return
+
+        pbar = tqdm(desc=f"{self.__class__.__name__} running", disable=self.disable_tqdm)
+        with ThreadPoolExecutor(max_workers=self.max_concurrency) as executor:
+            for future in as_completed([
+                executor.submit(self.process_rewriting, document, b[0], b[1])
+                for b in inference_blocks
+            ]):
+                future.result()  # Raise exceptions if any occurred
+                pbar.update(1)
+
+        pbar.close()
+
+    def get_block_lines(self, block: Block, document: Document) -> Tuple[list, list]:
+        text_lines = block.contained_blocks(document, (BlockTypes.Line,))
+        extracted_lines = [line.formatted_text(document) for line in text_lines]
+        return text_lines, extracted_lines
+
+    def process_rewriting(self, document: Document, page: PageGroup, block: Block):
+        SpanClass = get_block_class(BlockTypes.Span)
+
+        text_lines, extracted_lines = self.get_block_lines(block, document)
+        prompt = (self.text_math_rewriting_prompt.replace("{extracted_lines}",
+                                                         json.dumps({"extracted_lines": extracted_lines}, indent=2))
+                                                    .replace("{input_line_count}", str(len(extracted_lines)))
+                  )
+
+        image = self.extract_image(document, block)
+        response = self.llm_service(prompt, image, block, LLMTextSchema)
+
+        if not response or "corrected_lines" not in response:
+            block.update_metadata(llm_error_count=1)
+            return
+
+        corrected_lines = response["corrected_lines"]
+        if not corrected_lines or len(corrected_lines) != len(extracted_lines):
+            block.update_metadata(llm_error_count=1)
+            return
+
+        for text_line, corrected_text in zip(text_lines, corrected_lines):
+            text_line.structure = []
+            corrected_spans = text_to_spans(corrected_text)
+
+            for span_idx, span in enumerate(corrected_spans):
+                if span_idx == len(corrected_spans) - 1:
+                    span['content'] += "\n"
+
+                span_block = page.add_full_block(
+                    SpanClass(
+                        polygon=text_line.polygon,
+                        text=span['content'],
+                        font='Unknown',
+                        font_weight=0,
+                        font_size=0,
+                        minimum_position=0,
+                        maximum_position=0,
+                        formats=[span['type']],
+                        url=span.get('url'),
+                        page_id=text_line.page_id,
+                        text_extraction_method="gemini",
+                    )
+                )
+                text_line.structure.append(span_block.id)
+
+class LLMTextSchema(BaseModel):
+    corrected_lines: List[str]
\ No newline at end of file
diff --git a/marker/processors/llm/llm_text.py b/marker/processors/llm/llm_text.py
index 15a6e5b1..ae23619f 100644
--- a/marker/processors/llm/llm_text.py
+++ b/marker/processors/llm/llm_text.py
@@ -38,7 +38,7 @@ class LLMTextProcessor(BaseLLMSimpleBlockProcessor):
     * Other inaccuracies:  If the image is handwritten then you may correct any spelling errors, or other discrepancies.
 5. Do not remove any formatting i.e bold, italics, math, superscripts, subscripts, etc from the extracted lines unless it is necessary to correct an error.  The formatting 
 6. The number of corrected lines in the output MUST equal the number of extracted lines provided in the input. Do not add or remove lines.  There are exactly {line_count} input lines.
-7. Output the corrected lines in JSON format with a "lines" field, as shown in the example below.  Each line should be in HTML format. Only use the math, br, a, i, b, sup, sub, and span tags.
+7. Output the corrected lines in JSON format, as shown in the example below.  Each line should be in HTML format. Only use the math, br, a, i, b, sup, sub, and span tags.
 8. You absolutely cannot remove any <a href='#...'>...</a> tags, those are extremely important for references and are coming directly from the document, you MUST always preserve them.
 
 **Example:**
@@ -88,6 +88,7 @@ def inference_blocks(self, document: Document) -> List[List[BlockData]]:
                         "block": block
                     })
 
+
         out_blocks = []
         for i in range(0, len(blocks), self.math_line_batch_size):
             batch = blocks[i:i + self.math_line_batch_size]
diff --git a/marker/renderers/__init__.py b/marker/renderers/__init__.py
index af4a31c5..2a8cbe77 100644
--- a/marker/renderers/__init__.py
+++ b/marker/renderers/__init__.py
@@ -40,6 +40,17 @@ def extract_image(self, document: Document, image_id, to_base64=False):
             cropped = base64.b64encode(image_buffer.getvalue()).decode(settings.OUTPUT_ENCODING)
         return cropped
 
+    @staticmethod
+    def merge_consecutive_math(html, tag="math"):
+        if not html:
+            return html
+        pattern = fr'-</{tag}>(\s*)<{tag}>'
+        html = re.sub(pattern, " ", html)
+
+        pattern = fr'-</{tag}>(\s*)<{tag} display="inline">'
+        html = re.sub(pattern, " ", html)
+        return html
+
     @staticmethod
     def merge_consecutive_tags(html, tag):
         if not html:
diff --git a/marker/renderers/html.py b/marker/renderers/html.py
index bf8501aa..afe76c2b 100644
--- a/marker/renderers/html.py
+++ b/marker/renderers/html.py
@@ -82,6 +82,7 @@ def extract_html(self, document, document_output, level=0):
         if level == 0:
             output = self.merge_consecutive_tags(output, 'b')
             output = self.merge_consecutive_tags(output, 'i')
+            output = self.merge_consecutive_math(output) # Merge consecutive inline math tags
             output = textwrap.dedent(f"""
             <!DOCTYPE html>
             <html>

From a3932edc0bb7fe5057c5269966b737ca40c2594f Mon Sep 17 00:00:00 2001
From: Vik Paruchuri <vik.paruchuri@gmail.com>
Date: Thu, 20 Feb 2025 21:03:09 -0500
Subject: [PATCH 10/46] handle line merging for inline math better

---
 marker/builders/line.py         | 26 ++++++----
 marker/converters/pdf.py        |  2 +
 marker/processors/debug.py      |  2 +-
 marker/processors/line_merge.py | 91 +++++++++++++++++++++++++++++++++
 marker/schema/text/line.py      |  8 +++
 5 files changed, 117 insertions(+), 12 deletions(-)
 create mode 100644 marker/processors/line_merge.py

diff --git a/marker/builders/line.py b/marker/builders/line.py
index ef0da08a..7fd70991 100644
--- a/marker/builders/line.py
+++ b/marker/builders/line.py
@@ -113,7 +113,7 @@ def __init__(self, detection_model: DetectionPredictor, inline_detection_model:
     def __call__(self, document: Document, provider: PdfProvider):
         # Disable Inline Detection for documents where layout model doesn't detect any equations
         # Also disable if we won't use the inline detections (if we aren't using the LLM or texify)
-        do_inline_math_detection = document.contained_blocks([BlockTypes.Equation]) and (self.texify_inline_spans or self.use_llm)
+        do_inline_math_detection = document.contained_blocks([BlockTypes.Equation, BlockTypes.TextInlineMath]) and (self.texify_inline_spans or self.use_llm)
         provider_lines, ocr_lines = self.get_all_lines(document, provider, do_inline_math_detection)
         self.merge_blocks(document, provider_lines, ocr_lines)
 
@@ -146,6 +146,7 @@ def get_detection_results(self, page_images: List[Image.Image], run_detection: L
                 batch_size=self.get_detection_batch_size()
             )
 
+        assert len(page_detection_results) == len(inline_detection_results) == sum(run_detection)
         detection_results = []
         inline_results = []
         idx = 0
@@ -220,7 +221,7 @@ def get_all_lines(self, document: Document, provider: PdfProvider, do_inline_mat
                 page_lines[document_page.page_id].extend(
                     self.merge_provider_lines_inline_math(
                         provider_lines,
-                        [b for _,b in math_detection_boxes],
+                        merged_detection_boxes,
                         image_size,
                         page_size
                     )
@@ -388,47 +389,49 @@ def add_math_span_format(self, provider_line):
     def merge_provider_lines_inline_math(
         self,
         provider_lines: List[ProviderOutput],
-        inline_math_lines: List[TextBox],
+        text_lines: List[TextBox],
         image_size,
         page_size
     ):
         # When provider lines is empty or no inline math detected, return provider lines
-        if not provider_lines or not inline_math_lines:
+        if not provider_lines or not text_lines:
             return provider_lines
 
         horizontal_provider_lines = [
             (j, provider_line) for j, provider_line in enumerate(provider_lines)
-            if provider_line.line.polygon.height < provider_line.line.polygon.width * 3 # Multiply to account for small blocks inside equations, but filter out big vertical lines
+            if provider_line.line.polygon.height < provider_line.line.polygon.width * 5 # Multiply to account for small blocks inside equations, but filter out big vertical lines
         ]
         provider_line_boxes = [p.line.polygon.bbox for _, p in horizontal_provider_lines]
-        math_line_boxes = [PolygonBox(polygon=m.polygon).rescale(image_size, page_size).bbox for m in inline_math_lines]
+        math_line_boxes = [PolygonBox(polygon=m.polygon).rescale(image_size, page_size).bbox for m in text_lines]
 
         overlaps = matrix_intersection_area(provider_line_boxes, math_line_boxes)
 
         # Find potential merges
         merge_lines = defaultdict(list)
         for i in range(len(provider_line_boxes)):
-            max_overlap_pct = np.max(overlaps[i]) / horizontal_provider_lines[i][1].line.polygon.area
+            max_overlap_pct = np.max(overlaps[i]) / max(1, horizontal_provider_lines[i][1].line.polygon.area)
             if max_overlap_pct <= self.line_inline_min_overlap_pct:
                 continue
 
             best_overlap = np.argmax(overlaps[i])
-            best_overlap_line = horizontal_provider_lines[best_overlap]
-
             merge_lines[best_overlap].append(i)
 
         # Handle the merging
         already_merged = set()
         potential_merges = set(chain.from_iterable(merge_lines.values()))
         out_provider_lines = [(i, p) for i, p in enumerate(provider_lines) if i not in potential_merges]
-        for merge_section in merge_lines.values():
+        for line_idx in merge_lines:
+            text_line = text_lines[line_idx]
+            merge_section = merge_lines[line_idx]
             merge_section = [m for m in merge_section if m not in already_merged]
             if len(merge_section) == 0:
                 continue
             elif len(merge_section) == 1:
                 line_idx = merge_section[0]
                 merged_line = provider_lines[line_idx]
-                self.add_math_span_format(merged_line)
+                # Only add math format to single lines if the detected line is math
+                if text_line.math:
+                    self.add_math_span_format(merged_line)
                 out_provider_lines.append((line_idx, merged_line))
                 already_merged.add(merge_section[0])
                 continue
@@ -443,6 +446,7 @@ def merge_provider_lines_inline_math(
                 else:
                     # Combine the spans of the provider line with the merged line
                     merged_line = merged_line.merge(provider_line)
+                    # Add math regardless, since we assume heavily broken lines are math lines
                     self.add_math_span_format(merged_line)
                 already_merged.add(idx) # Prevent double merging
             out_provider_lines.append((min_idx, merged_line))
diff --git a/marker/converters/pdf.py b/marker/converters/pdf.py
index c324ca2d..b8ce8b55 100644
--- a/marker/converters/pdf.py
+++ b/marker/converters/pdf.py
@@ -1,5 +1,6 @@
 import os
 
+from marker.processors.line_merge import LineMergeProcessor
 from marker.processors.llm.llm_inlinemath import LLMInlineMathProcessor
 
 os.environ["TOKENIZERS_PARALLELISM"] = "false"  # disables a tokenizers warning
@@ -64,6 +65,7 @@ class PdfConverter(BaseConverter):
     ] = False
     default_processors: Tuple[BaseProcessor, ...] = (
         OrderProcessor,
+        LineMergeProcessor,
         BlockquoteProcessor,
         CodeProcessor,
         DocumentTOCProcessor,
diff --git a/marker/processors/debug.py b/marker/processors/debug.py
index e44a611b..0972c845 100644
--- a/marker/processors/debug.py
+++ b/marker/processors/debug.py
@@ -81,7 +81,7 @@ def draw_pdf_debug_images(self, document: Document):
                     span_bboxes.append(bbox)
 
             self.render_on_image(line_bboxes, png_image, color="blue", draw_bbox=True, label_font_size=24)
-            self.render_on_image(span_bboxes, png_image, color="green", draw_bbox=True, label_font_size=24)
+            #self.render_on_image(span_bboxes, png_image, color="green", draw_bbox=True, label_font_size=24)
 
             png_image = self.render_layout_boxes(page, png_image)
 
diff --git a/marker/processors/line_merge.py b/marker/processors/line_merge.py
new file mode 100644
index 00000000..3f3012c2
--- /dev/null
+++ b/marker/processors/line_merge.py
@@ -0,0 +1,91 @@
+from typing import Annotated
+
+from marker.processors import BaseProcessor
+from marker.schema import BlockTypes
+from marker.schema.document import Document
+from marker.schema.text import Line
+from marker.util import matrix_intersection_area
+
+
+class LineMergeProcessor(BaseProcessor):
+    """
+    A processor for merging inline math lines.
+    """
+    block_types = (BlockTypes.Text, BlockTypes.TextInlineMath)
+    min_merge_pct: Annotated[
+        float,
+        "The minimum percentage of intersection area to consider merging."
+    ] = .02
+    min_merge_ydist: Annotated[
+        float,
+        "The minimum y distance between lines to consider merging."
+    ] = 5
+    intersection_pct_threshold: Annotated[
+        float,
+        "The total amount of intersection area concentrated in the max intersection block."
+    ] = .9
+
+    def __init__(self, config):
+        super().__init__(config)
+
+    def __call__(self, document: Document):
+        for page in document.pages:
+            for block in page.contained_blocks(document, self.block_types):
+                if block.structure is None:
+                    continue
+
+                if not len(block.structure) >= 2:  # Skip single lines
+                    continue
+
+                lines = block.contained_blocks(document, (BlockTypes.Line,))
+                line_bboxes = [l.polygon.bbox for l in lines]
+                intersections = matrix_intersection_area(line_bboxes, line_bboxes)
+
+                merges = []
+                merge = []
+                for i in range(len(line_bboxes) - 1):
+                    next_idx = i + 1
+                    intersection_val = intersections[i, next_idx]
+                    intersection_pct = intersection_val / max(1, lines[i].polygon.area)
+                    intersection_row = intersections[i]
+                    intersection_row[i] = 0 # Zero out the current idx
+                    max_intersection_idx = intersection_row.argmax()
+                    total_intersection = max(1, sum(intersection_row))
+                    max_intersection = intersection_row[max_intersection_idx]
+
+
+                    if all([
+                        max_intersection_idx == next_idx, # The next line is the max intersection line
+                        intersection_pct >= self.min_merge_pct,
+                        abs(lines[i].polygon.y_start - lines[next_idx].polygon.y_start) <= self.min_merge_ydist,
+                        abs(lines[i].polygon.y_end - lines[next_idx].polygon.y_end) <= self.min_merge_ydist,
+                        max_intersection / total_intersection >= self.intersection_pct_threshold
+                    ]):
+                        merge.append(i)
+                    else:
+                        merges.append(merge)
+                        merge = []
+
+                if merge:
+                    merges.append(merge)
+
+                merges = [m for m in merges if len(m) > 1]
+                merged = set()
+                for merge in merges:
+                    merge = [m for m in merge if m not in merged]
+                    if len(merge) < 2:
+                        continue
+
+                    line: Line = lines[merge[0]]
+                    merged.add(merge[0])
+                    for idx in merge[1:]:
+                        other_line: Line = lines[idx]
+                        line.merge(other_line)
+                        block.structure.remove(other_line.id)
+                        merged.add(idx)
+
+                    # It is probably math if we are merging provider lines like this
+                    if not line.formats:
+                        line.formats = ["math"]
+                    elif "math" not in line.formats:
+                        line.formats.append("math")
diff --git a/marker/schema/text/line.py b/marker/schema/text/line.py
index 9e8a0141..4b3bb861 100644
--- a/marker/schema/text/line.py
+++ b/marker/schema/text/line.py
@@ -91,3 +91,11 @@ def render(self, document, parent_structure, section_hierarchy=None):
             children=[],
             section_hierarchy=section_hierarchy
         )
+
+    def merge(self, other: "Line"):
+        self.polygon = self.polygon.merge([other.polygon])
+        self.structure = self.structure + other.structure
+        if self.formats is None:
+            self.formats = other.formats
+        elif other.formats is not None:
+            self.formats.extend(other.formats)

From 87670d7725f05b7dffb4c9613c7e45ebc3ee49a2 Mon Sep 17 00:00:00 2001
From: Vik Paruchuri <vik.paruchuri@gmail.com>
Date: Fri, 21 Feb 2025 11:21:18 -0500
Subject: [PATCH 11/46] Update inline math line merges

---
 marker/builders/line.py                 |  30 +++++-
 marker/processors/debug.py              |   7 +-
 marker/processors/line_merge.py         |  14 ++-
 marker/processors/llm/llm_inlinemath.py |   4 +
 marker/schema/blocks/base.py            |   2 +-
 marker/services/claude.py               | 133 ++++++++++++++++++++++++
 poetry.lock                             | 130 ++++++++++++++++++++++-
 pyproject.toml                          |   1 +
 8 files changed, 306 insertions(+), 15 deletions(-)
 create mode 100644 marker/services/claude.py

diff --git a/marker/builders/line.py b/marker/builders/line.py
index 7fd70991..0e3ebe95 100644
--- a/marker/builders/line.py
+++ b/marker/builders/line.py
@@ -113,7 +113,11 @@ def __init__(self, detection_model: DetectionPredictor, inline_detection_model:
     def __call__(self, document: Document, provider: PdfProvider):
         # Disable Inline Detection for documents where layout model doesn't detect any equations
         # Also disable if we won't use the inline detections (if we aren't using the LLM or texify)
-        do_inline_math_detection = document.contained_blocks([BlockTypes.Equation, BlockTypes.TextInlineMath]) and (self.texify_inline_spans or self.use_llm)
+        do_inline_math_detection = all([
+            len(document.contained_blocks([BlockTypes.Equation, BlockTypes.TextInlineMath])) > 0,
+            (self.texify_inline_spans or self.use_llm)
+        ])
+
         provider_lines, ocr_lines = self.get_all_lines(document, provider, do_inline_math_detection)
         self.merge_blocks(document, provider_lines, ocr_lines)
 
@@ -186,7 +190,7 @@ def get_all_lines(self, document: Document, provider: PdfProvider, do_inline_mat
         if sum(layout_good) > len(document.pages) * self.min_document_ocr_threshold:
             layout_good = [True] * len(document.pages)
 
-        run_detection = [not good or do_inline_math_detection for good in layout_good]
+        run_detection = [(not good or do_inline_math_detection) for good in layout_good]
         page_images = [page.get_image(highres=False, remove_blocks=self.ocr_remove_blocks) for page, good in zip(document.pages, run_detection) if good]
 
         # Note: run_detection is longer than page_images, since it has a value for each page, not just good ones
@@ -416,13 +420,29 @@ def merge_provider_lines_inline_math(
             best_overlap = np.argmax(overlaps[i])
             merge_lines[best_overlap].append(i)
 
+        # Filter to get rid of detected lines that include multiple provider lines
+        filtered_merge_lines = {}
+        for line_idx in merge_lines:
+            first_line = horizontal_provider_lines[merge_lines[line_idx][0]][1].line.polygon
+            all_close = all([
+                (
+                    abs(horizontal_provider_lines[ml][1].line.polygon.y_start - first_line.y_start) < self.inline_math_line_vertical_merge_threshold
+                    or
+                    abs(horizontal_provider_lines[ml][1].line.polygon.y_end - first_line.y_end) < self.inline_math_line_vertical_merge_threshold
+                )
+                for ml in
+                merge_lines[line_idx]
+            ])
+            if all_close:
+                filtered_merge_lines[line_idx] = merge_lines[line_idx]
+
         # Handle the merging
         already_merged = set()
-        potential_merges = set(chain.from_iterable(merge_lines.values()))
+        potential_merges = set(chain.from_iterable(filtered_merge_lines.values()))
         out_provider_lines = [(i, p) for i, p in enumerate(provider_lines) if i not in potential_merges]
-        for line_idx in merge_lines:
+        for line_idx in filtered_merge_lines:
             text_line = text_lines[line_idx]
-            merge_section = merge_lines[line_idx]
+            merge_section = filtered_merge_lines[line_idx]
             merge_section = [m for m in merge_section if m not in already_merged]
             if len(merge_section) == 0:
                 continue
diff --git a/marker/processors/debug.py b/marker/processors/debug.py
index 0972c845..78c28a16 100644
--- a/marker/processors/debug.py
+++ b/marker/processors/debug.py
@@ -72,15 +72,20 @@ def draw_pdf_debug_images(self, document: Document):
 
             line_bboxes = []
             span_bboxes = []
+            line_ids = []
             for child in page.children:
+                # Skip any blocks that have been removed
+                if child.removed:
+                    continue
                 if child.block_type == BlockTypes.Line:
                     bbox = child.polygon.rescale(page.polygon.size, png_image.size).bbox
                     line_bboxes.append(bbox)
+                    line_ids.append(child.block_id)
                 elif child.block_type == BlockTypes.Span:
                     bbox = child.polygon.rescale(page.polygon.size, png_image.size).bbox
                     span_bboxes.append(bbox)
 
-            self.render_on_image(line_bboxes, png_image, color="blue", draw_bbox=True, label_font_size=24)
+            self.render_on_image(line_bboxes, png_image, color="blue", draw_bbox=True, label_font_size=24, labels=[str(i) for i in line_ids])
             #self.render_on_image(span_bboxes, png_image, color="green", draw_bbox=True, label_font_size=24)
 
             png_image = self.render_layout_boxes(page, png_image)
diff --git a/marker/processors/line_merge.py b/marker/processors/line_merge.py
index 3f3012c2..2ba2b42b 100644
--- a/marker/processors/line_merge.py
+++ b/marker/processors/line_merge.py
@@ -23,7 +23,7 @@ class LineMergeProcessor(BaseProcessor):
     intersection_pct_threshold: Annotated[
         float,
         "The total amount of intersection area concentrated in the max intersection block."
-    ] = .9
+    ] = .7
 
     def __init__(self, config):
         super().__init__(config)
@@ -38,7 +38,8 @@ def __call__(self, document: Document):
                     continue
 
                 lines = block.contained_blocks(document, (BlockTypes.Line,))
-                line_bboxes = [l.polygon.bbox for l in lines]
+                lines = [l for l in lines if l.polygon.width * 5 > l.polygon.height]  # Skip vertical lines
+                line_bboxes = [l.polygon.expand(self.min_merge_pct, 0).bbox for l in lines] # Expand horizontally
                 intersections = matrix_intersection_area(line_bboxes, line_bboxes)
 
                 merges = []
@@ -49,6 +50,10 @@ def __call__(self, document: Document):
                     intersection_pct = intersection_val / max(1, lines[i].polygon.area)
                     intersection_row = intersections[i]
                     intersection_row[i] = 0 # Zero out the current idx
+
+                    # Zero out previous merge segments
+                    for m in merge:
+                        intersection_row[m] = 0
                     max_intersection_idx = intersection_row.argmax()
                     total_intersection = max(1, sum(intersection_row))
                     max_intersection = intersection_row[max_intersection_idx]
@@ -61,7 +66,9 @@ def __call__(self, document: Document):
                         abs(lines[i].polygon.y_end - lines[next_idx].polygon.y_end) <= self.min_merge_ydist,
                         max_intersection / total_intersection >= self.intersection_pct_threshold
                     ]):
-                        merge.append(i)
+                        if not merge:
+                            merge.append(i)
+                        merge.append(next_idx)
                     else:
                         merges.append(merge)
                         merge = []
@@ -82,6 +89,7 @@ def __call__(self, document: Document):
                         other_line: Line = lines[idx]
                         line.merge(other_line)
                         block.structure.remove(other_line.id)
+                        other_line.removed = True # Mark line as removed
                         merged.add(idx)
 
                     # It is probably math if we are merging provider lines like this
diff --git a/marker/processors/llm/llm_inlinemath.py b/marker/processors/llm/llm_inlinemath.py
index 426e2a92..eba3601d 100644
--- a/marker/processors/llm/llm_inlinemath.py
+++ b/marker/processors/llm/llm_inlinemath.py
@@ -83,11 +83,14 @@ def rewrite_blocks(self, document: Document):
         if not self.redo_inline_math:
             return
 
+        # Get inline math blocks
         inline_blocks = [
             (page, block)
             for page in document.pages
             for block in page.contained_blocks(document, self.block_types)
         ]
+
+        # Get other blocks with detected math in them
         detected_blocks = [
             (page, block)
             for page in document.pages
@@ -95,6 +98,7 @@ def rewrite_blocks(self, document: Document):
             if any([b.formats and "math" in b.formats for b in block.contained_blocks(document, (BlockTypes.Line,))])
         ]
         inference_blocks = inline_blocks + detected_blocks
+
         # Don't show progress if there are no blocks to process
         total_blocks = len(inference_blocks)
         if total_blocks == 0:
diff --git a/marker/schema/blocks/base.py b/marker/schema/blocks/base.py
index 21fe0468..f68ea799 100644
--- a/marker/schema/blocks/base.py
+++ b/marker/schema/blocks/base.py
@@ -215,7 +215,7 @@ def contained_blocks(self, document: Document, block_types: Sequence[BlockTypes]
         blocks = []
         for block_id in self.structure:
             block = document.get_block(block_id)
-            if block_types is None or block.block_type in block_types:
+            if (block_types is None or block.block_type in block_types) and not block.removed:
                 blocks.append(block)
             blocks += block.contained_blocks(document, block_types)
         return blocks
diff --git a/marker/services/claude.py b/marker/services/claude.py
new file mode 100644
index 00000000..0ae29340
--- /dev/null
+++ b/marker/services/claude.py
@@ -0,0 +1,133 @@
+import base64
+import json
+import time
+from io import BytesIO
+from typing import List, Annotated, Union, T
+
+import PIL
+from PIL import Image
+import anthropic
+from anthropic import APIError, APIConnectionError, APITimeoutError, RateLimitError
+from pydantic import BaseModel
+
+from marker.schema.blocks import Block
+from marker.services import BaseService
+
+class ClaudeService(BaseService):
+    claude_model_name: Annotated[
+        str,
+        "The name of the Google model to use for the service."
+    ] = "claude-3-5-sonnet-20241022"
+    claude_api_key: Annotated[
+        str,
+        "The Claude API key to use for the service."
+    ] = None
+    max_claude_tokens: Annotated[
+        int,
+        "The maximum number of tokens to use for a single Claude request."
+    ] = 4096
+
+
+    def img_to_base64(self, img: PIL.Image.Image):
+        image_bytes = BytesIO()
+        img.save(image_bytes, format="WEBP")
+        return base64.b64encode(image_bytes.getvalue()).decode('utf-8')
+
+    def prepare_images(self, images: Union[Image.Image, List[Image.Image]]) -> List[dict]:
+        if isinstance(images, Image.Image):
+            images = [images]
+
+        return [
+            {
+                "type": "image",
+                "source": {
+                    "type": "base64",
+                    "media_type": "image/webp",
+                    "data": self.img_to_base64(img)
+                }
+            }
+            for img in images
+        ]
+
+    def validate_response(self, response_text: str, schema: type[T]) -> T:
+        try:
+            # Try to parse as JSON first
+            data = json.loads(response_text)
+            return schema.parse_obj(data)
+        except json.JSONDecodeError:
+            # If not JSON, try to parse the raw text into the schema
+            return schema.parse_raw(response_text)
+
+    def get_client(self):
+        return anthropic.Anthropic(
+            api_key=self.claude_api_key,
+        )
+
+    def __call__(
+            self,
+            prompt: str,
+            image: PIL.Image.Image | List[PIL.Image.Image],
+            block: Block,
+            response_schema: type[BaseModel],
+            max_retries: int | None = None,
+            timeout: int | None = None
+    ):
+        if max_retries is None:
+            max_retries = self.max_retries
+
+        if timeout is None:
+            timeout = self.timeout
+
+        if not isinstance(image, list):
+            image = [image]
+
+        schema_example = response_schema.model_json_schema()
+        system_prompt = f"""
+Follow the instructions given by the user prompt.  You must provide your response in JSON format matching this schema:
+
+{json.dumps(schema_example, indent=2)}
+""".strip()
+
+        client = self.get_client()
+        image_data = self.prepare_images(image)
+
+        messages = [
+            {
+                "role": "system",
+                "content": system_prompt
+            },
+            {
+                "role": "user",
+                "content": [
+                    {
+                        "type": "text",
+                        "text": prompt
+                    },
+                    *image_data
+                ]
+            }
+        ]
+
+        tries = 0
+        while tries < max_retries:
+            try:
+                response = client.messages.create(
+                    model=self.claude_model_name,
+                    max_tokens=self.max_claude_tokens,
+                    messages=messages,
+                    timeout=timeout
+                )
+                # Extract and validate response
+                response_text = response.content[0].text
+                return self.validate_response(response_text, response_schema)
+            except RateLimitError as e:
+                # Rate limit exceeded
+                tries += 1
+                wait_time = tries * 3
+                print(f"Rate limit error: {e}. Retrying in {wait_time} seconds... (Attempt {tries}/{max_retries})")
+                time.sleep(wait_time)
+            except Exception as e:
+                print(e)
+                break
+
+        return {}
\ No newline at end of file
diff --git a/poetry.lock b/poetry.lock
index d9191345..aa0b9332 100644
--- a/poetry.lock
+++ b/poetry.lock
@@ -163,6 +163,30 @@ files = [
     {file = "annotated_types-0.7.0.tar.gz", hash = "sha256:aff07c09a53a08bc8cfccb9c85b05f1aa9a2a6f23728d790723543408344ce89"},
 ]
 
+[[package]]
+name = "anthropic"
+version = "0.46.0"
+description = "The official Python library for the anthropic API"
+optional = false
+python-versions = ">=3.8"
+files = [
+    {file = "anthropic-0.46.0-py3-none-any.whl", hash = "sha256:1445ec9be78d2de7ea51b4d5acd3574e414aea97ef903d0ecbb57bec806aaa49"},
+    {file = "anthropic-0.46.0.tar.gz", hash = "sha256:eac3d43271d02321a57c3ca68aca84c3d58873e8e72d1433288adee2d46b745b"},
+]
+
+[package.dependencies]
+anyio = ">=3.5.0,<5"
+distro = ">=1.7.0,<2"
+httpx = ">=0.23.0,<1"
+jiter = ">=0.4.0,<1"
+pydantic = ">=1.9.0,<3"
+sniffio = "*"
+typing-extensions = ">=4.10,<5"
+
+[package.extras]
+bedrock = ["boto3 (>=1.28.57)", "botocore (>=1.31.57)"]
+vertex = ["google-auth (>=2,<3)"]
+
 [[package]]
 name = "anyio"
 version = "4.8.0"
@@ -777,6 +801,17 @@ files = [
     {file = "Distance-0.1.3.tar.gz", hash = "sha256:60807584f5b6003f5c521aa73f39f51f631de3be5cccc5a1d67166fcbf0d4551"},
 ]
 
+[[package]]
+name = "distro"
+version = "1.9.0"
+description = "Distro - an OS platform information API"
+optional = false
+python-versions = ">=3.6"
+files = [
+    {file = "distro-1.9.0-py3-none-any.whl", hash = "sha256:7bffd925d65168f85027d8da9af6bddab658135b840670a223589bc0c8ef02b2"},
+    {file = "distro-1.9.0.tar.gz", hash = "sha256:2fa77c6fd8940f116ee1d6b94a2f90b13b5ea8d019b98bc8bafdcabcdd9bdbed"},
+]
+
 [[package]]
 name = "exceptiongroup"
 version = "1.2.2"
@@ -1450,6 +1485,91 @@ MarkupSafe = ">=2.0"
 [package.extras]
 i18n = ["Babel (>=2.7)"]
 
+[[package]]
+name = "jiter"
+version = "0.8.2"
+description = "Fast iterable JSON parser."
+optional = false
+python-versions = ">=3.8"
+files = [
+    {file = "jiter-0.8.2-cp310-cp310-macosx_10_12_x86_64.whl", hash = "sha256:ca8577f6a413abe29b079bc30f907894d7eb07a865c4df69475e868d73e71c7b"},
+    {file = "jiter-0.8.2-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:b25bd626bde7fb51534190c7e3cb97cee89ee76b76d7585580e22f34f5e3f393"},
+    {file = "jiter-0.8.2-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:d5c826a221851a8dc028eb6d7d6429ba03184fa3c7e83ae01cd6d3bd1d4bd17d"},
+    {file = "jiter-0.8.2-cp310-cp310-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:d35c864c2dff13dfd79fb070fc4fc6235d7b9b359efe340e1261deb21b9fcb66"},
+    {file = "jiter-0.8.2-cp310-cp310-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:f557c55bc2b7676e74d39d19bcb8775ca295c7a028246175d6a8b431e70835e5"},
+    {file = "jiter-0.8.2-cp310-cp310-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:580ccf358539153db147e40751a0b41688a5ceb275e6f3e93d91c9467f42b2e3"},
+    {file = "jiter-0.8.2-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:af102d3372e917cffce49b521e4c32c497515119dc7bd8a75665e90a718bbf08"},
+    {file = "jiter-0.8.2-cp310-cp310-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:cadcc978f82397d515bb2683fc0d50103acff2a180552654bb92d6045dec2c49"},
+    {file = "jiter-0.8.2-cp310-cp310-musllinux_1_1_aarch64.whl", hash = "sha256:ba5bdf56969cad2019d4e8ffd3f879b5fdc792624129741d3d83fc832fef8c7d"},
+    {file = "jiter-0.8.2-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:3b94a33a241bee9e34b8481cdcaa3d5c2116f575e0226e421bed3f7a6ea71cff"},
+    {file = "jiter-0.8.2-cp310-cp310-win32.whl", hash = "sha256:6e5337bf454abddd91bd048ce0dca5134056fc99ca0205258766db35d0a2ea43"},
+    {file = "jiter-0.8.2-cp310-cp310-win_amd64.whl", hash = "sha256:4a9220497ca0cb1fe94e3f334f65b9b5102a0b8147646118f020d8ce1de70105"},
+    {file = "jiter-0.8.2-cp311-cp311-macosx_10_12_x86_64.whl", hash = "sha256:2dd61c5afc88a4fda7d8b2cf03ae5947c6ac7516d32b7a15bf4b49569a5c076b"},
+    {file = "jiter-0.8.2-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:a6c710d657c8d1d2adbbb5c0b0c6bfcec28fd35bd6b5f016395f9ac43e878a15"},
+    {file = "jiter-0.8.2-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:a9584de0cd306072635fe4b89742bf26feae858a0683b399ad0c2509011b9dc0"},
+    {file = "jiter-0.8.2-cp311-cp311-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:5a90a923338531b7970abb063cfc087eebae6ef8ec8139762007188f6bc69a9f"},
+    {file = "jiter-0.8.2-cp311-cp311-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:d21974d246ed0181558087cd9f76e84e8321091ebfb3a93d4c341479a736f099"},
+    {file = "jiter-0.8.2-cp311-cp311-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:32475a42b2ea7b344069dc1e81445cfc00b9d0e3ca837f0523072432332e9f74"},
+    {file = "jiter-0.8.2-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:8b9931fd36ee513c26b5bf08c940b0ac875de175341cbdd4fa3be109f0492586"},
+    {file = "jiter-0.8.2-cp311-cp311-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:ce0820f4a3a59ddced7fce696d86a096d5cc48d32a4183483a17671a61edfddc"},
+    {file = "jiter-0.8.2-cp311-cp311-musllinux_1_1_aarch64.whl", hash = "sha256:8ffc86ae5e3e6a93765d49d1ab47b6075a9c978a2b3b80f0f32628f39caa0c88"},
+    {file = "jiter-0.8.2-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:5127dc1abd809431172bc3fbe8168d6b90556a30bb10acd5ded41c3cfd6f43b6"},
+    {file = "jiter-0.8.2-cp311-cp311-win32.whl", hash = "sha256:66227a2c7b575720c1871c8800d3a0122bb8ee94edb43a5685aa9aceb2782d44"},
+    {file = "jiter-0.8.2-cp311-cp311-win_amd64.whl", hash = "sha256:cde031d8413842a1e7501e9129b8e676e62a657f8ec8166e18a70d94d4682855"},
+    {file = "jiter-0.8.2-cp312-cp312-macosx_10_12_x86_64.whl", hash = "sha256:e6ec2be506e7d6f9527dae9ff4b7f54e68ea44a0ef6b098256ddf895218a2f8f"},
+    {file = "jiter-0.8.2-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:76e324da7b5da060287c54f2fabd3db5f76468006c811831f051942bf68c9d44"},
+    {file = "jiter-0.8.2-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:180a8aea058f7535d1c84183c0362c710f4750bef66630c05f40c93c2b152a0f"},
+    {file = "jiter-0.8.2-cp312-cp312-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:025337859077b41548bdcbabe38698bcd93cfe10b06ff66617a48ff92c9aec60"},
+    {file = "jiter-0.8.2-cp312-cp312-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:ecff0dc14f409599bbcafa7e470c00b80f17abc14d1405d38ab02e4b42e55b57"},
+    {file = "jiter-0.8.2-cp312-cp312-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:ffd9fee7d0775ebaba131f7ca2e2d83839a62ad65e8e02fe2bd8fc975cedeb9e"},
+    {file = "jiter-0.8.2-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:14601dcac4889e0a1c75ccf6a0e4baf70dbc75041e51bcf8d0e9274519df6887"},
+    {file = "jiter-0.8.2-cp312-cp312-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:92249669925bc1c54fcd2ec73f70f2c1d6a817928480ee1c65af5f6b81cdf12d"},
+    {file = "jiter-0.8.2-cp312-cp312-musllinux_1_1_aarch64.whl", hash = "sha256:e725edd0929fa79f8349ab4ec7f81c714df51dc4e991539a578e5018fa4a7152"},
+    {file = "jiter-0.8.2-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:bf55846c7b7a680eebaf9c3c48d630e1bf51bdf76c68a5f654b8524335b0ad29"},
+    {file = "jiter-0.8.2-cp312-cp312-win32.whl", hash = "sha256:7efe4853ecd3d6110301665a5178b9856be7e2a9485f49d91aa4d737ad2ae49e"},
+    {file = "jiter-0.8.2-cp312-cp312-win_amd64.whl", hash = "sha256:83c0efd80b29695058d0fd2fa8a556490dbce9804eac3e281f373bbc99045f6c"},
+    {file = "jiter-0.8.2-cp313-cp313-macosx_10_12_x86_64.whl", hash = "sha256:ca1f08b8e43dc3bd0594c992fb1fd2f7ce87f7bf0d44358198d6da8034afdf84"},
+    {file = "jiter-0.8.2-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:5672a86d55416ccd214c778efccf3266b84f87b89063b582167d803246354be4"},
+    {file = "jiter-0.8.2-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:58dc9bc9767a1101f4e5e22db1b652161a225874d66f0e5cb8e2c7d1c438b587"},
+    {file = "jiter-0.8.2-cp313-cp313-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:37b2998606d6dadbb5ccda959a33d6a5e853252d921fec1792fc902351bb4e2c"},
+    {file = "jiter-0.8.2-cp313-cp313-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:4ab9a87f3784eb0e098f84a32670cfe4a79cb6512fd8f42ae3d0709f06405d18"},
+    {file = "jiter-0.8.2-cp313-cp313-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:79aec8172b9e3c6d05fd4b219d5de1ac616bd8da934107325a6c0d0e866a21b6"},
+    {file = "jiter-0.8.2-cp313-cp313-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:711e408732d4e9a0208008e5892c2966b485c783cd2d9a681f3eb147cf36c7ef"},
+    {file = "jiter-0.8.2-cp313-cp313-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:653cf462db4e8c41995e33d865965e79641ef45369d8a11f54cd30888b7e6ff1"},
+    {file = "jiter-0.8.2-cp313-cp313-musllinux_1_1_aarch64.whl", hash = "sha256:9c63eaef32b7bebac8ebebf4dabebdbc6769a09c127294db6babee38e9f405b9"},
+    {file = "jiter-0.8.2-cp313-cp313-musllinux_1_1_x86_64.whl", hash = "sha256:eb21aaa9a200d0a80dacc7a81038d2e476ffe473ffdd9c91eb745d623561de05"},
+    {file = "jiter-0.8.2-cp313-cp313-win32.whl", hash = "sha256:789361ed945d8d42850f919342a8665d2dc79e7e44ca1c97cc786966a21f627a"},
+    {file = "jiter-0.8.2-cp313-cp313-win_amd64.whl", hash = "sha256:ab7f43235d71e03b941c1630f4b6e3055d46b6cb8728a17663eaac9d8e83a865"},
+    {file = "jiter-0.8.2-cp313-cp313t-macosx_11_0_arm64.whl", hash = "sha256:b426f72cd77da3fec300ed3bc990895e2dd6b49e3bfe6c438592a3ba660e41ca"},
+    {file = "jiter-0.8.2-cp313-cp313t-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:b2dd880785088ff2ad21ffee205e58a8c1ddabc63612444ae41e5e4b321b39c0"},
+    {file = "jiter-0.8.2-cp313-cp313t-win_amd64.whl", hash = "sha256:3ac9f578c46f22405ff7f8b1f5848fb753cc4b8377fbec8470a7dc3997ca7566"},
+    {file = "jiter-0.8.2-cp38-cp38-macosx_10_12_x86_64.whl", hash = "sha256:9e1fa156ee9454642adb7e7234a383884452532bc9d53d5af2d18d98ada1d79c"},
+    {file = "jiter-0.8.2-cp38-cp38-macosx_11_0_arm64.whl", hash = "sha256:0cf5dfa9956d96ff2efb0f8e9c7d055904012c952539a774305aaaf3abdf3d6c"},
+    {file = "jiter-0.8.2-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:e52bf98c7e727dd44f7c4acb980cb988448faeafed8433c867888268899b298b"},
+    {file = "jiter-0.8.2-cp38-cp38-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:a2ecaa3c23e7a7cf86d00eda3390c232f4d533cd9ddea4b04f5d0644faf642c5"},
+    {file = "jiter-0.8.2-cp38-cp38-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:08d4c92bf480e19fc3f2717c9ce2aa31dceaa9163839a311424b6862252c943e"},
+    {file = "jiter-0.8.2-cp38-cp38-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:99d9a1eded738299ba8e106c6779ce5c3893cffa0e32e4485d680588adae6db8"},
+    {file = "jiter-0.8.2-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:d20be8b7f606df096e08b0b1b4a3c6f0515e8dac296881fe7461dfa0fb5ec817"},
+    {file = "jiter-0.8.2-cp38-cp38-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:d33f94615fcaf872f7fd8cd98ac3b429e435c77619777e8a449d9d27e01134d1"},
+    {file = "jiter-0.8.2-cp38-cp38-musllinux_1_1_aarch64.whl", hash = "sha256:317b25e98a35ffec5c67efe56a4e9970852632c810d35b34ecdd70cc0e47b3b6"},
+    {file = "jiter-0.8.2-cp38-cp38-musllinux_1_1_x86_64.whl", hash = "sha256:fc9043259ee430ecd71d178fccabd8c332a3bf1e81e50cae43cc2b28d19e4cb7"},
+    {file = "jiter-0.8.2-cp38-cp38-win32.whl", hash = "sha256:fc5adda618205bd4678b146612ce44c3cbfdee9697951f2c0ffdef1f26d72b63"},
+    {file = "jiter-0.8.2-cp38-cp38-win_amd64.whl", hash = "sha256:cd646c827b4f85ef4a78e4e58f4f5854fae0caf3db91b59f0d73731448a970c6"},
+    {file = "jiter-0.8.2-cp39-cp39-macosx_10_12_x86_64.whl", hash = "sha256:e41e75344acef3fc59ba4765df29f107f309ca9e8eace5baacabd9217e52a5ee"},
+    {file = "jiter-0.8.2-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:7f22b16b35d5c1df9dfd58843ab2cd25e6bf15191f5a236bed177afade507bfc"},
+    {file = "jiter-0.8.2-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:f7200b8f7619d36aa51c803fd52020a2dfbea36ffec1b5e22cab11fd34d95a6d"},
+    {file = "jiter-0.8.2-cp39-cp39-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:70bf4c43652cc294040dbb62256c83c8718370c8b93dd93d934b9a7bf6c4f53c"},
+    {file = "jiter-0.8.2-cp39-cp39-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:f9d471356dc16f84ed48768b8ee79f29514295c7295cb41e1133ec0b2b8d637d"},
+    {file = "jiter-0.8.2-cp39-cp39-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:859e8eb3507894093d01929e12e267f83b1d5f6221099d3ec976f0c995cb6bd9"},
+    {file = "jiter-0.8.2-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:eaa58399c01db555346647a907b4ef6d4f584b123943be6ed5588c3f2359c9f4"},
+    {file = "jiter-0.8.2-cp39-cp39-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:8f2d5ed877f089862f4c7aacf3a542627c1496f972a34d0474ce85ee7d939c27"},
+    {file = "jiter-0.8.2-cp39-cp39-musllinux_1_1_aarch64.whl", hash = "sha256:03c9df035d4f8d647f8c210ddc2ae0728387275340668fb30d2421e17d9a0841"},
+    {file = "jiter-0.8.2-cp39-cp39-musllinux_1_1_x86_64.whl", hash = "sha256:8bd2a824d08d8977bb2794ea2682f898ad3d8837932e3a74937e93d62ecbb637"},
+    {file = "jiter-0.8.2-cp39-cp39-win32.whl", hash = "sha256:ca29b6371ebc40e496995c94b988a101b9fbbed48a51190a4461fcb0a68b4a36"},
+    {file = "jiter-0.8.2-cp39-cp39-win_amd64.whl", hash = "sha256:1c0dfbd1be3cbefc7510102370d86e35d1d53e5a93d48519688b1bf0f761160a"},
+    {file = "jiter-0.8.2.tar.gz", hash = "sha256:cd73d3e740666d0e639f678adb176fad25c1bcbdae88d8d7b857e1783bb4212d"},
+]
+
 [[package]]
 name = "joblib"
 version = "1.4.2"
@@ -2691,10 +2811,10 @@ files = [
 
 [package.dependencies]
 numpy = [
-    {version = ">=1.26.0", markers = "python_version >= \"3.12\""},
-    {version = ">=1.23.5", markers = "python_version >= \"3.11\" and python_version < \"3.12\""},
     {version = ">=1.21.4", markers = "python_version >= \"3.10\" and platform_system == \"Darwin\" and python_version < \"3.11\""},
     {version = ">=1.21.2", markers = "platform_system != \"Darwin\" and python_version >= \"3.10\" and python_version < \"3.11\""},
+    {version = ">=1.23.5", markers = "python_version >= \"3.11\" and python_version < \"3.12\""},
+    {version = ">=1.26.0", markers = "python_version >= \"3.12\""},
 ]
 
 [[package]]
@@ -2772,9 +2892,9 @@ files = [
 
 [package.dependencies]
 numpy = [
-    {version = ">=1.26.0", markers = "python_version >= \"3.12\""},
-    {version = ">=1.23.2", markers = "python_version == \"3.11\""},
     {version = ">=1.22.4", markers = "python_version < \"3.11\""},
+    {version = ">=1.23.2", markers = "python_version == \"3.11\""},
+    {version = ">=1.26.0", markers = "python_version >= \"3.12\""},
 ]
 python-dateutil = ">=2.8.2"
 pytz = ">=2020.1"
@@ -5453,4 +5573,4 @@ propcache = ">=0.2.0"
 [metadata]
 lock-version = "2.0"
 python-versions = "^3.10"
-content-hash = "04afa6e305c60db8d9f5d304d67f6c51e0415cee9b3c7d1171750f7dd787135b"
+content-hash = "664bbdbf1226d34671e66f96973588ea266c9e0aefaace7d67da5a7cd00f0eec"
diff --git a/pyproject.toml b/pyproject.toml
index bb47df3e..4d151f46 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -35,6 +35,7 @@ markdown2 = "^2.5.2"
 filetype = "^1.2.0"
 scikit-learn = "^1.6.1"
 google-genai = "^1.0.0"
+anthropic = "^0.46.0"
 
 [tool.poetry.group.dev.dependencies]
 jupyter = "^1.0.0"

From 84bbd99f201fe384415bbf02fd8410f24e230a18 Mon Sep 17 00:00:00 2001
From: Vik Paruchuri <vik.paruchuri@gmail.com>
Date: Fri, 21 Feb 2025 12:41:24 -0500
Subject: [PATCH 12/46] Misc fixes

---
 marker/builders/line.py                 |  2 +-
 marker/processors/llm/llm_equation.py   |  2 +-
 marker/processors/llm/llm_inlinemath.py | 18 +++++++++++++-----
 marker/services/claude.py               | 24 +++++++++++++++---------
 pyproject.toml                          |  2 +-
 5 files changed, 31 insertions(+), 17 deletions(-)

diff --git a/marker/builders/line.py b/marker/builders/line.py
index 0e3ebe95..4133c374 100644
--- a/marker/builders/line.py
+++ b/marker/builders/line.py
@@ -111,7 +111,7 @@ def __init__(self, detection_model: DetectionPredictor, inline_detection_model:
         self.ocr_error_model = ocr_error_model
 
     def __call__(self, document: Document, provider: PdfProvider):
-        # Disable Inline Detection for documents where layout model doesn't detect any equations
+        # Disable inline detection for documents where layout model doesn't detect any equations
         # Also disable if we won't use the inline detections (if we aren't using the LLM or texify)
         do_inline_math_detection = all([
             len(document.contained_blocks([BlockTypes.Equation, BlockTypes.TextInlineMath])) > 0,
diff --git a/marker/processors/llm/llm_equation.py b/marker/processors/llm/llm_equation.py
index 0529fc85..e1568842 100644
--- a/marker/processors/llm/llm_equation.py
+++ b/marker/processors/llm/llm_equation.py
@@ -12,7 +12,7 @@ class LLMEquationProcessor(BaseLLMSimpleBlockProcessor):
     min_equation_height: Annotated[
         float,
         "The minimum ratio between equation height and page height to consider for processing.",
-     ] = 0.08
+     ] = 0.06
     image_expansion_ratio: Annotated[
         float,
         "The ratio to expand the image by when cropping.",
diff --git a/marker/processors/llm/llm_inlinemath.py b/marker/processors/llm/llm_inlinemath.py
index eba3601d..dd429566 100644
--- a/marker/processors/llm/llm_inlinemath.py
+++ b/marker/processors/llm/llm_inlinemath.py
@@ -32,12 +32,12 @@ class LLMInlineMathProcessor(BaseLLMComplexBlockProcessor):
 1. Carefully examine the provided text block image .
 2. Analyze the extracted lines.
 3. For each extracted line, compare it to the corresponding line in the image.
-4. Correct any errors in the extracted line, including:
+4. If there are no errors in any of the extracted lines, output "No corrections needed".
+5. For each extracted line, correct any errors, including:
     * Inline math: Ensure all mathematical expressions are correctly formatted and rendered.  Surround them with <math>...</math> tags.
     * Formatting: Maintain consistent formatting with the text block image, including spacing, indentation, and special characters.
     * Other inaccuracies:  If the image is handwritten then you may correct any spelling errors, or other discrepancies.
-5. Do not remove any formatting i.e bold, italics, math, superscripts, subscripts, etc from the extracted lines unless it is necessary to correct an error.
-6. Ensure that inline math is properly with inline math tags, like <math>...</math>.
+6. Do not remove any formatting i.e bold, italics, math, superscripts, subscripts, etc from the extracted lines unless it is necessary to correct an error.
 7. The number of corrected lines in the output MUST equal the number of extracted lines provided in the input. Do not add or remove lines.
 8. Output the corrected lines in JSON format, as shown in the example below.  Each line should be in HTML format. Only use the math, br, a, i, b, sup, sub, and span tags.
 9. You absolutely cannot remove any <a href='#...'>...</a> tags, those are extremely important for references and are coming directly from the document, you MUST always preserve them.
@@ -94,7 +94,7 @@ def rewrite_blocks(self, document: Document):
         detected_blocks = [
             (page, block)
             for page in document.pages
-            for block in page.contained_blocks(document, (BlockTypes.Text, BlockTypes.Caption, BlockTypes.SectionHeader))
+            for block in page.contained_blocks(document, (BlockTypes.Text, BlockTypes.Caption, BlockTypes.SectionHeader, BlockTypes.Footnote))
             if any([b.formats and "math" in b.formats for b in block.contained_blocks(document, (BlockTypes.Line,))])
         ]
         inference_blocks = inline_blocks + detected_blocks
@@ -137,7 +137,15 @@ def process_rewriting(self, document: Document, page: PageGroup, block: Block):
             return
 
         corrected_lines = response["corrected_lines"]
-        if not corrected_lines or len(corrected_lines) != len(extracted_lines):
+        if not corrected_lines:
+            block.update_metadata(llm_error_count=1)
+            return
+
+        # Block is fine
+        if "no corrections needed" in str(corrected_lines).lower():
+            return
+
+        if len(corrected_lines) != len(extracted_lines):
             block.update_metadata(llm_error_count=1)
             return
 
diff --git a/marker/services/claude.py b/marker/services/claude.py
index 0ae29340..9899ab11 100644
--- a/marker/services/claude.py
+++ b/marker/services/claude.py
@@ -7,7 +7,7 @@
 import PIL
 from PIL import Image
 import anthropic
-from anthropic import APIError, APIConnectionError, APITimeoutError, RateLimitError
+from anthropic import RateLimitError
 from pydantic import BaseModel
 
 from marker.schema.blocks import Block
@@ -50,13 +50,20 @@ def prepare_images(self, images: Union[Image.Image, List[Image.Image]]) -> List[
         ]
 
     def validate_response(self, response_text: str, schema: type[T]) -> T:
+        response_text = response_text.strip()
+        if response_text.startswith("```json"):
+            response_text = response_text[7:]
+        if response_text.endswith("```"):
+            response_text = response_text[:-3]
         try:
             # Try to parse as JSON first
-            data = json.loads(response_text)
-            return schema.parse_obj(data)
+            out_schema = schema.model_validate_json(response_text)
+            out_json = out_schema.model_dump()
+            return out_json
         except json.JSONDecodeError:
             # If not JSON, try to parse the raw text into the schema
-            return schema.parse_raw(response_text)
+            out_schema = schema.model_validate_strings(response_text)
+            return out_schema.model_dump()
 
     def get_client(self):
         return anthropic.Anthropic(
@@ -86,24 +93,22 @@ def __call__(
 Follow the instructions given by the user prompt.  You must provide your response in JSON format matching this schema:
 
 {json.dumps(schema_example, indent=2)}
+
+Respond only with the JSON schema, nothing else.  Do not include ```json``` or any other formatting.
 """.strip()
 
         client = self.get_client()
         image_data = self.prepare_images(image)
 
         messages = [
-            {
-                "role": "system",
-                "content": system_prompt
-            },
             {
                 "role": "user",
                 "content": [
+                    *image_data,
                     {
                         "type": "text",
                         "text": prompt
                     },
-                    *image_data
                 ]
             }
         ]
@@ -112,6 +117,7 @@ def __call__(
         while tries < max_retries:
             try:
                 response = client.messages.create(
+                    system=system_prompt,
                     model=self.claude_model_name,
                     max_tokens=self.max_claude_tokens,
                     messages=messages,
diff --git a/pyproject.toml b/pyproject.toml
index 4d151f46..9d8a5577 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "marker-pdf"
-version = "1.5.5"
+version = "1.5.6"
 description = "Convert PDF to markdown with high speed and accuracy."
 authors = ["Vik Paruchuri <github@vikas.sh>"]
 readme = "README.md"

From 8ebafc9c4d8872d24e7c74620ee8bd52923d78ce Mon Sep 17 00:00:00 2001
From: Vik Paruchuri <vik.paruchuri@gmail.com>
Date: Fri, 21 Feb 2025 14:24:46 -0500
Subject: [PATCH 13/46] Fix JSON parsing

---
 marker/services/claude.py | 13 +++++++++----
 1 file changed, 9 insertions(+), 4 deletions(-)

diff --git a/marker/services/claude.py b/marker/services/claude.py
index 9899ab11..1442a682 100644
--- a/marker/services/claude.py
+++ b/marker/services/claude.py
@@ -55,15 +55,20 @@ def validate_response(self, response_text: str, schema: type[T]) -> T:
             response_text = response_text[7:]
         if response_text.endswith("```"):
             response_text = response_text[:-3]
+
         try:
             # Try to parse as JSON first
             out_schema = schema.model_validate_json(response_text)
             out_json = out_schema.model_dump()
             return out_json
-        except json.JSONDecodeError:
-            # If not JSON, try to parse the raw text into the schema
-            out_schema = schema.model_validate_strings(response_text)
-            return out_schema.model_dump()
+        except Exception as e:
+            try:
+                # Re-parse with fixed escapes
+                escaped_str = response_text.replace('\\', '\\\\')
+                out_schema = schema.model_validate_json(escaped_str)
+                return out_schema.model_dump()
+            except Exception as e:
+                return
 
     def get_client(self):
         return anthropic.Anthropic(

From d07122dd161b109dbbbf80289db976fa5f7edb75 Mon Sep 17 00:00:00 2001
From: Vik Paruchuri <vik.paruchuri@gmail.com>
Date: Fri, 21 Feb 2025 14:57:14 -0500
Subject: [PATCH 14/46] Test line merging

---
 marker/converters/pdf.py                 |  6 ++---
 marker/services/claude.py                |  2 +-
 tests/builders/test_inline_math_lines.py | 31 ++++++++++++++++++++++++
 3 files changed, 34 insertions(+), 5 deletions(-)
 create mode 100644 tests/builders/test_inline_math_lines.py

diff --git a/marker/converters/pdf.py b/marker/converters/pdf.py
index b8ce8b55..abab04ed 100644
--- a/marker/converters/pdf.py
+++ b/marker/converters/pdf.py
@@ -1,8 +1,4 @@
 import os
-
-from marker.processors.line_merge import LineMergeProcessor
-from marker.processors.llm.llm_inlinemath import LLMInlineMathProcessor
-
 os.environ["TOKENIZERS_PARALLELISM"] = "false"  # disables a tokenizers warning
 
 from collections import defaultdict
@@ -46,6 +42,8 @@
 from marker.processors.llm.llm_handwriting import LLMHandwritingProcessor
 from marker.processors.order import OrderProcessor
 from marker.services.gemini import GoogleGeminiService
+from marker.processors.line_merge import LineMergeProcessor
+from marker.processors.llm.llm_inlinemath import LLMInlineMathProcessor
 
 
 class PdfConverter(BaseConverter):
diff --git a/marker/services/claude.py b/marker/services/claude.py
index 1442a682..aeef7a2b 100644
--- a/marker/services/claude.py
+++ b/marker/services/claude.py
@@ -99,7 +99,7 @@ def __call__(
 
 {json.dumps(schema_example, indent=2)}
 
-Respond only with the JSON schema, nothing else.  Do not include ```json``` or any other formatting.
+Respond only with the JSON schema, nothing else.  Do not include ```json, ```,  or any other formatting.
 """.strip()
 
         client = self.get_client()
diff --git a/tests/builders/test_inline_math_lines.py b/tests/builders/test_inline_math_lines.py
new file mode 100644
index 00000000..a718c3b6
--- /dev/null
+++ b/tests/builders/test_inline_math_lines.py
@@ -0,0 +1,31 @@
+import pytest
+
+from marker.processors.line_merge import LineMergeProcessor
+from marker.schema import BlockTypes
+
+@pytest.mark.config({"page_range": [1]})
+def test_inline_box_nomerging(pdf_document, config):
+    first_page = pdf_document.pages[0]
+    block = pdf_document.get_block(first_page.structure[1]) # First inline math block
+    line_count = len(block.contained_blocks(pdf_document, (BlockTypes.Line,)))
+    assert line_count == 46
+
+    merger = LineMergeProcessor(config)
+    merger(pdf_document)
+
+    line_count = len(block.contained_blocks(pdf_document, (BlockTypes.Line,)))
+    assert line_count == 45
+
+
+@pytest.mark.config({"page_range": [1], "use_llm": True})
+def test_inline_box_merging(pdf_document, config):
+    first_page = pdf_document.pages[0]
+    block = pdf_document.get_block(first_page.structure[1]) # First inline math block
+    line_count = len(block.contained_blocks(pdf_document, (BlockTypes.Line,)))
+    assert line_count == 21
+
+    merger = LineMergeProcessor(config)
+    merger(pdf_document)
+
+    line_count = len(block.contained_blocks(pdf_document, (BlockTypes.Line,)))
+    assert line_count == 21
\ No newline at end of file

From 82e632f527f04616d8d16b2d6b17d535137cfc8a Mon Sep 17 00:00:00 2001
From: Vik Paruchuri <vik.paruchuri@gmail.com>
Date: Fri, 21 Feb 2025 17:18:23 -0500
Subject: [PATCH 15/46] Fix HTML

---
 marker/converters/__init__.py   |  5 +++-
 marker/processors/debug.py      | 23 ++--------------
 marker/providers/__init__.py    | 21 +++++++++++++++
 marker/providers/document.py    |  2 +-
 marker/providers/epub.py        |  2 +-
 marker/providers/html.py        |  3 +++
 marker/providers/powerpoint.py  |  2 +-
 marker/providers/spreadsheet.py |  2 +-
 marker/scripts/file_to_s3.py    | 47 +++++++++++++++++++++++++++++++++
 marker/settings.py              |  3 +++
 marker/util.py                  | 14 +++++++++-
 11 files changed, 97 insertions(+), 27 deletions(-)
 create mode 100644 marker/scripts/file_to_s3.py

diff --git a/marker/converters/__init__.py b/marker/converters/__init__.py
index 9cd553b9..8357a499 100644
--- a/marker/converters/__init__.py
+++ b/marker/converters/__init__.py
@@ -6,7 +6,7 @@
 from marker.processors import BaseProcessor
 from marker.processors.llm import BaseLLMSimpleBlockProcessor
 from marker.processors.llm.llm_meta import LLMSimpleBlockMetaProcessor
-from marker.util import assign_config
+from marker.util import assign_config, download_font
 
 
 class BaseConverter:
@@ -15,6 +15,9 @@ def __init__(self, config: Optional[BaseModel | dict] = None):
         self.config = config
         self.llm_service = None
 
+        # Download render font, needed for some providers
+        download_font()
+
     def __call__(self, *args, **kwargs):
         raise NotImplementedError
 
diff --git a/marker/processors/debug.py b/marker/processors/debug.py
index e44a611b..8cfd6a96 100644
--- a/marker/processors/debug.py
+++ b/marker/processors/debug.py
@@ -2,7 +2,6 @@
 import os
 from typing import Annotated
 
-import requests
 from PIL import Image, ImageDraw, ImageFont
 
 from marker.processors import BaseProcessor
@@ -36,14 +35,7 @@ class DebugProcessor(BaseProcessor):
         bool,
         "Whether to dump block debug data.",
     ] = False
-    render_font: Annotated[
-        str,
-        "The path to the font to use for rendering debug images.",
-    ] = os.path.join(settings.FONT_DIR, "GoNotoCurrent-Regular.ttf")
-    font_dl_path: Annotated[
-        str,
-        "The path to download the font from.",
-    ] = "/~https://github.com/satbyy/go-noto-universal/releases/download/v7.0"
+
 
     def __call__(self, document: Document):
         # Remove extension from doc name
@@ -146,17 +138,6 @@ def dump_block_debug_data(self, document: Document):
         with open(debug_file, "w+") as f:
             json.dump(debug_data, f)
 
-    def get_font_path(self) -> str:
-        if not os.path.exists(self.render_font):
-            os.makedirs(os.path.dirname(self.render_font), exist_ok=True)
-            font_dl_path = f"{self.font_dl_path}/{os.path.basename(self.render_font)}"
-            with requests.get(font_dl_path, stream=True) as r, open(self.render_font, 'wb') as f:
-                r.raise_for_status()
-                for chunk in r.iter_content(chunk_size=8192):
-                    f.write(chunk)
-
-        return self.render_font
-
     def get_text_size(self, text, font):
         im = Image.new(mode="P", size=(0, 0))
         draw = ImageDraw.Draw(im)
@@ -165,7 +146,7 @@ def get_text_size(self, text, font):
 
     def render_on_image(self, bboxes, image, labels=None, label_offset=1, label_font_size=10, color: str | list = 'red', draw_bbox=True):
         draw = ImageDraw.Draw(image)
-        font_path = self.get_font_path()
+        font_path = settings.FONT_PATH
         label_font = ImageFont.truetype(font_path, label_font_size)
 
         for i, bbox in enumerate(bboxes):
diff --git a/marker/providers/__init__.py b/marker/providers/__init__.py
index b87092b5..85454067 100644
--- a/marker/providers/__init__.py
+++ b/marker/providers/__init__.py
@@ -5,11 +5,14 @@
 from pydantic import BaseModel
 
 from pdftext.schema import Reference
+from weasyprint import CSS
+from weasyprint.text.fonts import FontConfiguration
 
 from marker.logger import configure_logging
 from marker.schema.polygon import PolygonBox
 from marker.schema.text import Span
 from marker.schema.text.line import Line
+from marker.settings import settings
 from marker.util import assign_config
 
 configure_logging()
@@ -69,3 +72,21 @@ def get_page_refs(self, idx: int) -> List[Reference]:
 
     def __enter__(self):
         return self
+
+    @staticmethod
+    def get_font_css():
+        font_config = FontConfiguration()
+        css = CSS(string=f'''
+            @font-face {{
+                font-family: GoNotoCurrent-Regular;
+                src: url({settings.FONT_PATH});
+                font-display: swap;
+            }}
+            body {{
+                font-family: {settings.FONT_NAME.split(".")[0]}, sans-serif;
+                font-variant-ligatures: none;
+                font-feature-settings: "liga" 0;
+                text-rendering: optimizeLegibility;
+            }}
+            ''', font_config=font_config)
+        return css
diff --git a/marker/providers/document.py b/marker/providers/document.py
index ddd97e54..58621d1e 100644
--- a/marker/providers/document.py
+++ b/marker/providers/document.py
@@ -77,7 +77,7 @@ def convert_docx_to_pdf(self, filepath: str):
             # We convert the HTML into a PDF
             HTML(string=self._preprocess_base64_images(html)).write_pdf(
                 self.temp_pdf_path,
-                stylesheets=[CSS(string=css)]
+                stylesheets=[CSS(string=css), self.get_font_css()]
             )
 
     @staticmethod
diff --git a/marker/providers/epub.py b/marker/providers/epub.py
index c546372c..d50ce99b 100644
--- a/marker/providers/epub.py
+++ b/marker/providers/epub.py
@@ -105,5 +105,5 @@ def convert_epub_to_pdf(self, filepath):
         # we convert the epub to HTML
         result = HTML(string=html_content, base_url=filepath).write_pdf(
             self.temp_pdf_path,
-            stylesheets=[CSS(string=full_style)]
+            stylesheets=[CSS(string=full_style), self.get_font_css()]
         )
diff --git a/marker/providers/html.py b/marker/providers/html.py
index 4fb91097..6b24e918 100644
--- a/marker/providers/html.py
+++ b/marker/providers/html.py
@@ -18,6 +18,7 @@ def __init__(self, filepath: str, config=None):
             raise RuntimeError(f"Failed to convert {filepath} to PDF: {e}")
 
         # Initialize the PDF provider with the temp pdf path
+        print(self.temp_pdf_path)
         super().__init__(self.temp_pdf_path, config)
 
     def __del__(self):
@@ -25,6 +26,8 @@ def __del__(self):
             os.remove(self.temp_pdf_path)
 
     def convert_html_to_pdf(self, filepath: str):
+        font_css = self.get_font_css()
         HTML(filename=filepath, encoding="utf-8").write_pdf(
             self.temp_pdf_path,
+            stylesheets=[font_css]
         )
diff --git a/marker/providers/powerpoint.py b/marker/providers/powerpoint.py
index 331076c5..71e56d26 100644
--- a/marker/providers/powerpoint.py
+++ b/marker/providers/powerpoint.py
@@ -104,7 +104,7 @@ def convert_pptx_to_pdf(self, filepath):
         # We convert the HTML into a PDF
         HTML(string=html).write_pdf(
             self.temp_pdf_path,
-            stylesheets=[CSS(string=css)]
+            stylesheets=[CSS(string=css), self.get_font_css()]
         )
 
     def _handle_group(self, group_shape) -> str:
diff --git a/marker/providers/spreadsheet.py b/marker/providers/spreadsheet.py
index a66204bb..9ec3b428 100644
--- a/marker/providers/spreadsheet.py
+++ b/marker/providers/spreadsheet.py
@@ -64,7 +64,7 @@ def convert_xlsx_to_pdf(self, filepath: str):
         # We convert the HTML into a PDF
         HTML(string=html).write_pdf(
             self.temp_pdf_path,
-            stylesheets=[CSS(string=css)]
+            stylesheets=[CSS(string=css), self.get_font_css()]
         )
 
     @staticmethod
diff --git a/marker/scripts/file_to_s3.py b/marker/scripts/file_to_s3.py
new file mode 100644
index 00000000..c2ca46ef
--- /dev/null
+++ b/marker/scripts/file_to_s3.py
@@ -0,0 +1,47 @@
+import json
+import shutil
+import datetime
+from pathlib import Path
+import boto3
+
+from huggingface_hub import snapshot_download
+
+import click
+
+S3_API_URL = "https://1afbe4656a6b40d982ab5e730a39f6b9.r2.cloudflarestorage.com"
+
+@click.command(help="Uploads files to an S3 bucket")
+@click.argument("filepath", type=str)
+@click.argument("s3_path", type=str)
+@click.option("--bucket_name", type=str, default="datalab")
+@click.option("--access_key_id", type=str, default="<access_key_id>")
+@click.option("--access_key_secret", type=str, default="<access_key_secret>")
+def main(filepath: str, s3_path: str, bucket_name: str, access_key_id: str, access_key_secret: str):
+    filepath = Path(filepath)
+    # Upload the files to S3
+    s3_client = boto3.client(
+        's3',
+        endpoint_url=S3_API_URL,
+        aws_access_key_id=access_key_id,
+        aws_secret_access_key=access_key_secret,
+        region_name="enam"
+    )
+
+    s3_key = f"{s3_path}/{filepath.name}"
+
+    try:
+        s3_client.upload_file(
+            str(filepath),
+            bucket_name,
+            s3_key
+        )
+    except Exception as e:
+        print(f"Error uploading {filepath}: {str(e)}")
+
+    print(f"Uploaded files to {s3_path}")
+
+if __name__ == "__main__":
+    main()
+
+
+
diff --git a/marker/settings.py b/marker/settings.py
index 2b1eda90..caa605c8 100644
--- a/marker/settings.py
+++ b/marker/settings.py
@@ -13,6 +13,9 @@ class Settings(BaseSettings):
     OUTPUT_DIR: str = os.path.join(BASE_DIR, "conversion_results")
     FONT_DIR: str = os.path.join(BASE_DIR, "static", "fonts")
     DEBUG_DATA_FOLDER: str = os.path.join(BASE_DIR, "debug_data")
+    ARTIFACT_URL: str = "https://models.datalab.to/artifacts"
+    FONT_NAME: str = "GoNotoCurrent-Regular.ttf"
+    FONT_PATH: str = os.path.join(FONT_DIR, FONT_NAME)
 
     # General
     OUTPUT_ENCODING: str = "utf-8"
diff --git a/marker/util.py b/marker/util.py
index c44d084a..4a90fa26 100644
--- a/marker/util.py
+++ b/marker/util.py
@@ -1,11 +1,14 @@
 import inspect
+import os
 from importlib import import_module
 from typing import List, Annotated
 
 import numpy as np
+import requests
 from pydantic import BaseModel
 
 from marker.schema.polygon import PolygonBox
+from marker.settings import settings
 
 
 def strings_to_classes(items: List[str]) -> List[type]:
@@ -131,4 +134,13 @@ def sort_text_lines(lines: List[PolygonBox], tolerance=1.25):
         sorted_group = sorted(group, key=lambda x: x.bbox[0])
         sorted_lines.extend(sorted_group)
 
-    return sorted_lines
\ No newline at end of file
+    return sorted_lines
+
+def download_font():
+    if not os.path.exists(settings.FONT_PATH):
+        os.makedirs(os.path.dirname(settings.FONT_PATH), exist_ok=True)
+        font_dl_path = f"{settings.ARTIFACT_URL}/{settings.FONT_NAME}"
+        with requests.get(font_dl_path, stream=True) as r, open(settings.FONT_PATH, 'wb') as f:
+            r.raise_for_status()
+            for chunk in r.iter_content(chunk_size=8192):
+                f.write(chunk)
\ No newline at end of file

From 35f4bd8e413b8523dac4005a5755d3fb53c28871 Mon Sep 17 00:00:00 2001
From: Vik Paruchuri <vik.paruchuri@gmail.com>
Date: Fri, 21 Feb 2025 17:34:09 -0500
Subject: [PATCH 16/46] Add more tests

---
 README.md                              |  5 +--
 tests/converters/test_pdf_converter.py | 48 ++++++++++++++++++++++++++
 2 files changed, 51 insertions(+), 2 deletions(-)

diff --git a/README.md b/README.md
index f774c8a4..2c16b411 100644
--- a/README.md
+++ b/README.md
@@ -1,8 +1,8 @@
 # Marker
 
-Marker converts PDFs and images to markdown, JSON, and HTML quickly and accurately.
+Marker converts documents to markdown, JSON, and HTML quickly and accurately.
 
-- Supports a range of documents in all languages
+- Converts PDF, image, PPTX, DOCX, XLSX, HTML, EPUB in all languages
 - Formats tables, forms, equations, inline math, links, references, and code blocks
 - Extracts and saves images
 - Removes headers/footers/other artifacts
@@ -320,6 +320,7 @@ When running with the `--use_llm` flag, you have a choice of services you can us
 - `Gemini` - this will use the Gemini developer API by default.  You'll need to pass `--gemini_api_key` to configuration.
 - `Google Vertex` - this will use vertex, which can be more reliable.  You'll need to pass `--vertex_project_id`.  To use it, set `--llm_service=marker.services.vertex.GoogleVertexService`.
 - `Ollama` - this will use local models.  You can configure `--ollama_base_url` and `--ollama_model`. To use it, set `--llm_service=marker.services.ollama.OllamaService`.
+- `Claude` - this will use the anthropic API.  You can configure `--claude_api_key`, and `--claude_model_name`.  To use it, set `--llm_service=marker.services.claude.ClaudeService`.
 
 These services may have additional optional configuration as well - you can see it by viewing the classes.
 
diff --git a/tests/converters/test_pdf_converter.py b/tests/converters/test_pdf_converter.py
index b1fc4068..0aa584a7 100644
--- a/tests/converters/test_pdf_converter.py
+++ b/tests/converters/test_pdf_converter.py
@@ -20,3 +20,51 @@ def test_pdf_converter(pdf_converter: PdfConverter, temp_doc):
     # Some assertions for line joining across columns
     assert "remain similar across a wide range of choices." in markdown  # pg: 2
     assert "a new scheme for designing more robust and efficient" in markdown  # pg: 8
+
+@pytest.mark.filename("manual.epub")
+@pytest.mark.config({"page_range": [0]})
+def test_epub_converter(pdf_converter: PdfConverter, temp_doc):
+    markdown_output: MarkdownOutput = pdf_converter(temp_doc.name)
+    markdown = markdown_output.markdown
+
+    # Basic assertions
+    assert "Simple Sabotage Field Manual" in markdown
+
+@pytest.mark.filename("single_sheet.xlsx")
+@pytest.mark.config({"page_range": [0]})
+def test_xlsx_converter(pdf_converter: PdfConverter, temp_doc):
+    markdown_output: MarkdownOutput = pdf_converter(temp_doc.name)
+    markdown = markdown_output.markdown
+
+    # Basic assertions
+    assert "four" in markdown
+
+
+@pytest.mark.filename("china.html")
+@pytest.mark.config({"page_range": [10]})
+def test_html_converter(pdf_converter: PdfConverter, temp_doc):
+    markdown_output: MarkdownOutput = pdf_converter(temp_doc.name)
+    markdown = markdown_output.markdown
+
+    # Basic assertions
+    assert "Beijing" in markdown
+
+
+@pytest.mark.filename("gatsby.docx")
+@pytest.mark.config({"page_range": [0]})
+def test_docx_converter(pdf_converter: PdfConverter, temp_doc):
+    markdown_output: MarkdownOutput = pdf_converter(temp_doc.name)
+    markdown = markdown_output.markdown
+
+    # Basic assertions
+    assert "The Decline of the American Dream in the 1920s" in markdown
+
+
+@pytest.mark.filename("lambda.pptx")
+@pytest.mark.config({"page_range": [0]})
+def test_pptx_converter(pdf_converter: PdfConverter, temp_doc):
+    markdown_output: MarkdownOutput = pdf_converter(temp_doc.name)
+    markdown = markdown_output.markdown
+
+    # Basic assertions
+    assert "Adam Doupé" in markdown

From 1d46f0e806116e6add4985c099b9757d0e145848 Mon Sep 17 00:00:00 2001
From: Vik Paruchuri <vik.paruchuri@gmail.com>
Date: Fri, 21 Feb 2025 18:19:00 -0500
Subject: [PATCH 17/46] Fix tests

---
 marker/processors/llm/llm_table.py         |  2 ++
 tests/converters/test_pdf_converter.py     |  2 +-
 tests/providers/test_document_providers.py | 17 +----------------
 3 files changed, 4 insertions(+), 17 deletions(-)

diff --git a/marker/processors/llm/llm_table.py b/marker/processors/llm/llm_table.py
index e0669332..6aeb611d 100644
--- a/marker/processors/llm/llm_table.py
+++ b/marker/processors/llm/llm_table.py
@@ -171,6 +171,8 @@ def get_cell_text(element, keep_tags=('br','i', 'b', 'span', 'math')) -> str:
     def parse_html_table(self, html_text: str, block: Block, page: PageGroup) -> List[TableCell]:
         soup = BeautifulSoup(html_text, 'html.parser')
         table = soup.find('table')
+        if not table:
+            return []
 
         # Initialize grid
         rows = table.find_all('tr')
diff --git a/tests/converters/test_pdf_converter.py b/tests/converters/test_pdf_converter.py
index 0aa584a7..06a8eaa0 100644
--- a/tests/converters/test_pdf_converter.py
+++ b/tests/converters/test_pdf_converter.py
@@ -47,7 +47,7 @@ def test_html_converter(pdf_converter: PdfConverter, temp_doc):
     markdown = markdown_output.markdown
 
     # Basic assertions
-    assert "Beijing" in markdown
+    assert "Republic of China" in markdown
 
 
 @pytest.mark.filename("gatsby.docx")
diff --git a/tests/providers/test_document_providers.py b/tests/providers/test_document_providers.py
index e9526051..8697d2ee 100644
--- a/tests/providers/test_document_providers.py
+++ b/tests/providers/test_document_providers.py
@@ -4,14 +4,11 @@
 @pytest.mark.config({"page_range": [0]})
 @pytest.mark.filename("lambda.pptx")
 def test_pptx_provider(doc_provider):
-    assert len(doc_provider) == 22
     assert doc_provider.get_images([0], 72)[0].size == (842, 596)
 
     page_lines = doc_provider.get_page_lines(0)
-    assert len(page_lines) == 26
 
     spans = page_lines[0].spans
-    assert len(spans) == 2
     assert spans[0].text == "Lambda Calculus"
 
     spans = page_lines[1].spans
@@ -21,53 +18,41 @@ def test_pptx_provider(doc_provider):
 @pytest.mark.config({"page_range": [0]})
 @pytest.mark.filename("manual.epub")
 def test_epub_provider(doc_provider):
-    assert len(doc_provider) == 20
     assert doc_provider.get_images([0], 72)[0].size == (596, 842)
 
     page_lines = doc_provider.get_page_lines(0)
-    assert len(page_lines) == 31
 
     spans = page_lines[0].spans
-    assert len(spans) == 2
-    assert spans[0].text == "The Project Gutenberg eBook of Simple Sabotage Field"
+    assert spans[0].text == "The Project Gutenberg eBook of Simple"
 
 
 @pytest.mark.config({"page_range": [0]})
 @pytest.mark.filename("china.html")
 def test_html_provider(doc_provider):
-    assert len(doc_provider) == 73
     assert doc_provider.get_images([0], 72)[0].size == (596, 842)
 
     page_lines = doc_provider.get_page_lines(0)
-    assert len(page_lines) == 55
 
     spans = page_lines[0].spans
-    assert len(spans) == 2
     assert spans[0].text == "Jump to content"
 
 @pytest.mark.config({"page_range": [0]})
 @pytest.mark.filename("gatsby.docx")
 def test_docx_provider(doc_provider):
-    assert len(doc_provider) == 2
     assert doc_provider.get_images([0], 72)[0].size == (596, 842)
 
     page_lines = doc_provider.get_page_lines(0)
-    assert len(page_lines) == 54
 
     spans = page_lines[0].spans
-    assert len(spans) == 2
     assert spans[0].text == "Themes"
 
 
 @pytest.mark.config({"page_range": [0]})
 @pytest.mark.filename("single_sheet.xlsx")
 def test_xlsx_provider(doc_provider):
-    assert len(doc_provider) == 1
     assert doc_provider.get_images([0], 72)[0].size == (842, 596)
 
     page_lines = doc_provider.get_page_lines(0)
-    assert len(page_lines) == 4
 
     spans = page_lines[0].spans
-    assert len(spans) == 2
     assert spans[0].text == "Sheet1"
\ No newline at end of file

From 72863e251a76dc89786116b3d96d047e6c5b3ffb Mon Sep 17 00:00:00 2001
From: Vicencio <vicenciomf@uc.cl>
Date: Sat, 22 Feb 2025 01:19:02 -0300
Subject: [PATCH 18/46] Fix utf-8 encoding for JSON config files

Fix character encoding issues when loading configuration files with non-ASCII
characters.
---
 marker/config/parser.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/marker/config/parser.py b/marker/config/parser.py
index 7d5a65ea..1bff73e3 100644
--- a/marker/config/parser.py
+++ b/marker/config/parser.py
@@ -65,7 +65,7 @@ def generate_config_dict(self) -> Dict[str, any]:
                 case "languages":
                     config["languages"] = v.split(",")
                 case "config_json":
-                    with open(v, "r") as f:
+                    with open(v, "r", encoding="utf-8") as f:
                         config.update(json.load(f))
                 case "disable_multiprocessing":
                     config["pdftext_workers"] = 1

From 237325fbb9cf02300a33f5e0cbf56800e4236ade Mon Sep 17 00:00:00 2001
From: Vik Paruchuri <vik.paruchuri@gmail.com>
Date: Mon, 24 Feb 2025 11:48:12 -0500
Subject: [PATCH 19/46] Inline math updates

---
 README.md                                     |   1 +
 .../markdown/multicolcnn/multicolcnn.md       |  98 ++++----
 .../switch_transformers/switch_trans.md       | 210 ++++++++++--------
 marker/processors/debug.py                    |   2 +-
 marker/processors/line_merge.py               | 157 +++++++------
 marker/renderers/markdown.py                  |   3 +-
 poetry.lock                                   | 192 ++++++++--------
 pyproject.toml                                |   2 +-
 8 files changed, 360 insertions(+), 305 deletions(-)

diff --git a/README.md b/README.md
index 2c16b411..e3cc5f19 100644
--- a/README.md
+++ b/README.md
@@ -95,6 +95,7 @@ Options:
 - `--output_format [markdown|json|html]`: Specify the format for the output results.
 - `--paginate_output`: Paginates the output, using `\n\n{PAGE_NUMBER}` followed by `-` * 48, then `\n\n` 
 - `--use_llm`: Uses an LLM to improve accuracy.  You must set your Gemini API key using the `GOOGLE_API_KEY` env var.
+- `--redo_inline_math`: If you want the highest quality inline math conversion, use this along with `--use_llm`.
 - `--disable_image_extraction`: Don't extract images from the PDF.  If you also specify `--use_llm`, then images will be replaced with a description.
 - `--page_range TEXT`: Specify which pages to process. Accepts comma-separated page numbers and ranges. Example: `--page_range "0,5-10,20"` will process pages 0, 5 through 10, and page 20.
 - `--force_ocr`: Force OCR processing on the entire document, even for pages that might contain extractable text.
diff --git a/data/examples/markdown/multicolcnn/multicolcnn.md b/data/examples/markdown/multicolcnn/multicolcnn.md
index c4a7dccf..a74fd464 100644
--- a/data/examples/markdown/multicolcnn/multicolcnn.md
+++ b/data/examples/markdown/multicolcnn/multicolcnn.md
@@ -50,11 +50,11 @@ In this paper, however, we aim to apply the dilated convolution method of [\[25\
 
 ### <span id="page-2-0"></span>3.1. Dilated Convolutions for Multicolumn Networks
 
-We propose the use of dilated convolutions as an attractive alternative to the architecture of the HydraCNN [\[18\]](#page-8-4), which seems to saturate in performance at 3 or more columns. We refer to our proposed network as the aggregated multicolumn dilated convolution network[1](#page-2-1) , henceforth shortened as the AMDCN. The architecture of the AMDCN is inspired by the multicolumn counting network of [\[28\]](#page-9-0). Extracting features from multiple scales is a good idea when attempting to perform perspective-free counting and increasing the convolution kernel size across columns is an efficient method of doing so. However, the number of parameters increases exponentially as larger kernels are used in these columns to extract features at larger scales. Therefore, we propose using dilated convolutions rather than larger kernels.
+We propose the use of dilated convolutions as an attractive alternative to the architecture of the HydraCNN [\[18\]](#page-8-4), which seems to saturate in performance at 3 or more columns. We refer to our proposed network as the aggregated multicolumn dilated convolution network1, henceforth shortened as the AMDCN. The architecture of the AMDCN is inspired by the multicolumn counting network of [\[28\]](#page-9-0). Extracting features from multiple scales is a good idea when attempting to perform perspective-free counting and increasing the convolution kernel size across columns is an efficient method of doing so. However, the number of parameters increases exponentially as larger kernels are used in these columns to extract features at larger scales. Therefore, we propose using dilated convolutions rather than larger kernels.
 
 Dilated convolutions, as discussed in [\[25\]](#page-8-5), allow for the exponential increase of the receptive field with a linear increase in the number of parameters with respect to each hidden layer.
 
-In a traditional 2D convolution, we define a real valued function $F: \mathbb{Z}^2 \rightarrow \mathbb{R}$, an input $\Omega_r = [-r, r]^2 \in \mathbb{Z}^2$, and a filter function $k: \Omega_r \rightarrow \mathbb{R}$. In this case, a convolution operation as defined in [\[25\]](#page-8-5) is given by
+In a traditional 2D convolution, we define a real valued function $F : \mathbb{Z}^2 \rightarrow \mathbb{R}$, an input $\Omega_r = [-r, r]^2 \in \mathbb{Z}^2$, and a filter function $k : \Omega_r \rightarrow \mathbb{R}$. In this case, a convolution operation as defined in [\[25\]](#page-8-5) is given by
 
 $$(F*k)(\mathbf{p}) = \sum_{\mathbf{s}+\mathbf{t}=\mathbf{p}} F(\mathbf{s})k(\mathbf{t}).\tag{1}$$
 
@@ -66,17 +66,17 @@ where l is the index of the current layer of the convolution.
 
 Using dilations to construct the aggregator in combination with the multicolumn idea will allow for the construction of a network with more than just 3 or 4 columns as in [\[28\]](#page-9-0) and [\[8\]](#page-8-13), because the aggregator should prevent the saturation of performance with increasing numbers of columns. Therefore the network will be able to extract useful features from more scales. We take advantage of dilations within the columns as well to provide large receptive fields with fewer parameters.
 
-Looking at more scales should allow for more accurate regression of the density map. However, because not all scales will be relevant, we extend the network beyond a simple $1 \times 1$ convolution after the merged columns. Instead, we construct a second part of the network, the aggregator, which sets our method apart from [\[28\]](#page-9-0), [\[8\]](#page-8-13), and other multicolumn networks. This aggregator is another series of dilated convolutions that should appropriately consolidate the multiscale information collected by the columns. This is a capability of dilated convolutions observed by [\[25\]](#page-8-5). While papers such as [\[28\]](#page-9-0) and [\[8\]](#page-8-13) have shown that multiple columns and dilated columns are useful in extracting multiscale information, we argue in this paper that the simple aggregator module built using dilated convolutions is able to effectively make use multiscale information from multiple columns. We show compelling evidence for these claims in Section [4.5.](#page-5-0)
+Looking at more scales should allow for more accurate regression of the density map. However, because not all scales will be relevant, we extend the network beyond a simple 1 × 1 convolution after the merged columns. Instead, we construct a second part of the network, the aggregator, which sets our method apart from [\[28\]](#page-9-0), [\[8\]](#page-8-13), and other multicolumn networks. This aggregator is another series of dilated convolutions that should appropriately consolidate the multiscale information collected by the columns. This is a capability of dilated convolutions observed by [\[25\]](#page-8-5). While papers such as [\[28\]](#page-9-0) and [\[8\]](#page-8-13) have shown that multiple columns and dilated columns are useful in extracting multiscale information, we argue in this paper that the simple aggregator module built using dilated convolutions is able to effectively make use multiscale information from multiple columns. We show compelling evidence for these claims in Section 4.5.
 
-The network as shown in Figure [1](#page-1-0) contains 5 columns. Note that dilations allow us to use more columns for counting than [\[28\]](#page-9-0) or [\[8\]](#page-8-13). Each column looks at a larger scale than the previous (the exact dilations can also be seen in Figure [1)](#page-1-0). There are 32 feature maps for each convolution, and all inputs are zero padded prior to each convolution in order to maintain the same data shape from input to output. That is, an image input to this network will result in a density map of the same dimensions. All activations in the specified network are ReLUs. Our input pixel values are floating point 32 bit values from 0 to 1. We center our inputs at 0 by subtracting the per channel mean from each channel. When
+The network as shown in Figure [1](#page-1-0) contains 5 columns. Note that dilations allow us to use more columns for counting than [\[28\]](#page-9-0) or [\[8\]](#page-8-13). Each column looks at a larger scale than the previous (the exact dilations can also be seen in Figure [1](#page-1-0)). There are 32 feature maps for each convolution, and all inputs are zero padded prior to each convolution in order to maintain the same data shape from input to output. That is, an image input to this network will result in a density map of the same dimensions. All activations in the specified network are ReLUs. Our input pixel values are floating point 32 bit values from 0 to 1. We center our inputs at 0 by subtracting the per channel mean from each channel. When
 
-<span id="page-2-1"></span><sup>1</sup> Implementation available on [/~https://github.com/](/~https://github.com/diptodip/counting) [diptodip/counting](/~https://github.com/diptodip/counting).
+<span id="page-2-1"></span>1 Implementation available on [/~https://github.com/](/~https://github.com/diptodip/counting) [diptodip/counting](/~https://github.com/diptodip/counting).
 
 training, we use a scaled mean absolute error for our loss function:
 
 $$L = \frac{1}{n} \sum_{i=1}^{n} |\hat{y}_i - \gamma y_i| \tag{3}$$
 
-where γ is the scale factor, yˆi is the prediction, yi is the true value, and n is the number of pixels. We use a scaled mean absolute error because the target values are so small that it is numerically unstable to regress to these values. At testing time, when retrieving the output density map from the network, we scale the pixel values by $\gamma^{-1}$ to obtain the correct value. This approach is more numerically stable and avoids having the network learn to output only zeros by weighting the nonzero values highly. For all our datasets, we set $\gamma = 255$.
+where $\n \gamma\n $ is the scale factor, $\n \hat{y}_i\n $ is the prediction, $\n y_i\n $ is the true value, and $n$ is the number of pixels. We use a scaled mean absolute error because the target values are so small that it is numerically unstable to regress to these values. At testing time, when retrieving the output density map from the network, we scale the pixel values by $\gamma^{-1}$ to obtain the correct value. This approach is more numerically stable and avoids having the network learn to output only zeros by weighting the nonzero values highly. For all our datasets, we set $\gamma = 255$.
 
 #### 3.2. Experiments
 
@@ -84,7 +84,7 @@ We evaluated the performance of dilated convolutions against various counting me
 
 We have observed that multicolumn dilations produce density maps (and therefore counts) that often have lower loss than those of HydraCNN [\[18\]](#page-8-4) and [\[28\]](#page-9-0). We measure density map regression loss via a scaled mean absolute error loss during training. We compare accuracy of the counts via mean absolute error for the crowd datasets and the GAME metric in the TRANCOS dataset as explained in Section [3.2.2.](#page-3-0) Beyond the comparison to HydraCNN, we will also compare to other recent convolutional counting methods, especially those of [\[21\]](#page-8-14), [\[24\]](#page-8-15), and [\[4\]](#page-8-16) where possible.
 
-For all datasets, we generally use patched input images and ground truth density maps produced by summing a Gaussian of a fixed size ($\sigma$) for each object for training. This size varies from dataset to dataset, but remains constant within a dataset with the exception of cases in which a perspective map is used. This is explained per dataset. All experiments were performed using Keras with the Adam optimizer [\[10\]](#page-8-18). The learning rates used are detailed per dataset. For testing, we also use patches that can either be directly pieced together or overlapped and averaged except in the case of UCF, for which we run our network on the full image.
+For all datasets, we generally use patched input images and ground truth density maps produced by summing a Gaussian of a fixed size ($σ$) for each object for training. This size varies from dataset to dataset, but remains constant within a dataset with the exception of cases in which a perspective map is used. This is explained per dataset. All experiments were performed using Keras with the Adam optimizer [\[10\]](#page-8-18). The learning rates used are detailed per dataset. For testing, we also use patches that can either be directly pieced together or overlapped and averaged except in the case of UCF, for which we run our network on the full image.
 
 Furthermore, we performed a set of experiments in which we varied the number of columns from 1 to 5 (simply by including or not including the columns as specified in Figure [1,](#page-1-0) starting with the smallest filter column and adding larger filter columns one by one). Essentially, the network is allowed to extract information at larger and larger scales in addition to the smaller scales as we include each column. We then performed the same set of experiments, varying the number of columns, but with the aggregator module removed. We perform these experiments on the original split of UCSD as specified in Section [3.2.3](#page-4-0) and [\[5\]](#page-8-17), the TRAN-COS dataset, and the WorldExpo dataset because these are relatively large and well defined datasets. We limit the number of epochs to 10 for all of these sets of experiments in order to control for the effect of learning time, and also compare all results using MAE for consistency. These experiments are key to determining the efficacy of the aggregator in effectively combining multiscale information and in providing evidence to support the use of multiple columns to extract multiscale information from images. We report the results of these ablation studies in Section [4.5.](#page-5-0)
 
@@ -92,23 +92,23 @@ Furthermore, we performed a set of experiments in which we varied the number of
 
 UCF is a particularly challenging crowd counting dataset. There are only 50 images in the whole dataset and they are all of varying sizes and from different scenes. The number of people also varies between images from less than 100 to the thousands. The average image has on the order of 1000 people. The difficulty is due to the combination of the very low number of images in the dataset and the fact that the images are all of varying scenes, making high quality generalization crucial. Furthermore, perspective effects are particularly noticeable for many images in this dataset. Despite this, there is no perspective information available for this dataset.
 
-We take 1600 random patches of size $150 \times 150$ for the training. For testing, we do not densely scan the image as in [\[18\]](#page-8-4) but instead test on the whole image. In order to standardize the image sizes, we pad each image out with zeros until all images are $1024 \times 1024$. We then suppress output in the regions where we added padding when testing. This provides a cleaner resulting density map for these large crowds. The ground truth density maps are produced by annotating each object with a Gaussian of $\sigma = 15$.
+We take 1600 random patches of size $150 \,\times\, 150$ for the training. For testing, we do not densely scan the image as in [\[18\]](#page-8-4) but instead test on the whole image. In order to standardize the image sizes, we pad each image out with zeros until all images are $1024 \times 1024$. We then suppress output in the regions where we added padding when testing. This provides a cleaner resulting density map for these large crowds. The ground truth density maps are produced by annotating each object with a Gaussian of $\sigma = 15$.
 
 #### <span id="page-3-0"></span>3.2.2 TRANCOS Traffic Counting
 
-TRANCOS is a traffic counting dataset that comes with its own metric [\[14\]](#page-8-1). This metric is known as $GAME$, which stands for Grid Average Mean absolute Error. $GAME$ splits a given density map into 4 L grids, or subarrays, and obtains a mean absolute error within each grid separately. The value of L is a parameter chosen by the user. These individual errors are summed to obtain the final error for a particular image. The intuition behind this metric is that it is desirable to penalize a density map whose overall count might match the ground truth, but whose shape does not match the ground truth [\[14\]](#page-8-1). More formally, we define
+TRANCOS is a traffic counting dataset that comes with its own metric [\[14\]](#page-8-1). This metric is known as $GAME$, which stands for Grid Average Mean absolute Error. $GAME $splits a given density map into $4^{L}$ grids, or subarrays, and obtains a mean absolute error within each grid separately. The value of $L$ is a parameter chosen by the user. These individual errors are summed to obtain the final error for a particular image. The intuition behind this metric is that it is desirable to penalize a density map whose overall count might match the ground truth, but whose shape does not match the ground truth [\[14\]](#page-8-1). More formally, we define
 
 $$GAME(L) = \frac{1}{N} \cdot \sum_{n=1}^{N} \left( \sum_{l=1}^{4^L} |e_n^l - t_n^l| \right) \qquad (4)$$
 
-where N refers to the number of images, L is the level parameter for $GAME$, $e_n^l$ is the predicted or estimated count in region l of image n and t l n is the ground truth count in region l of image n [\[14\]](#page-8-1).
+where $N$ refers to the number of images, $L$ is the level parameter for $GAME$, $e_{n}^{l}$ is the predicted or estimated count in region $l$ of image $n$ and $t_{n}^{l}$ is the ground truth count in region $l$ of image $n$ [\[14\]](#page-8-1).
 
-For training this dataset, we take 1600 randomly sampled patches of size 80 × 80. For testing this dataset, we take 80 × 80 non-overlapping patches which we can stitch back together into the full-sized 640 × 480 images. We trained the AMDCN network with density maps produced with a Gaussian of $\sigma$ = 15 as specified in [\[18\]](#page-8-4).
+For training this dataset, we take 1600 randomly sampled patches of size 80 × 80. For testing this dataset, we take 80 × 80 non-overlapping patches which we can stitch back together into the full-sized 640 × 480 images. We trained the AMDCN network with density maps produced with a Gaussian of $σ$ = 15 as specified in [\[18\]](#page-8-4).
 
 #### <span id="page-4-0"></span>3.2.3 UCSD Crowd Counting
 
 The UCSD crowd counting dataset consists of frames of video of a sidewalk. There are relatively few people in view at any given time (approximately 25 on average). Furthermore, because the dataset comes from a video, there are many nearly identical images in the dataset. For this dataset, there have been two different ways to split the data into train and test sets. Therefore, we report results using both methods of splitting the data. The first method consists of four different splits: maximal, downscale, upscale, and minimal. Minimal is particularly challenging as the train set contains only 10 images. Moreover, upscale appears to be the easiest for the majority of methods [\[18\]](#page-8-4). The second method of splitting this data is much more succinct, leaving 1200 images in the testing set and 800 images in the training set [\[28\]](#page-9-0). This split comes from the original paper, so we call it the original split [\[5\]](#page-8-17).
 
-For this dataset, each object is annotated with a 2D Gaussian of covariance $\Sigma = 8 \cdot 1_{2\times2}$. The ground truth map is produced by summing these. When we make use of the perspective maps provided, we divide Σ by the perspective map value at that pixel $\mathbf{x}$, represented by $M(\mathbf{x})$. The provided perspective map for UCSD contains both a horizontal and vertical direction so we take the square root of the provided combined value. For training, we take 1600 random 79 × 119 pixel patches and for testing, we split each test image up into quadrants (which have dimension 79 × 119). There are two different ways to split the dataset into training and testing sets. We have experimented on the split that gave [\[18\]](#page-8-4) the best results as well as the split used in [\[28\]](#page-9-0).
+For this dataset, each object is annotated with a 2D Gaussian of covariance $\Sigma = 8 \cdot 1_{2\times2}$. The ground truth map is produced by summing these. When we make use of the perspective maps provided, we divide $\Sigma$ by the perspective map value at that pixel x, represented by $M(x)$. The provided perspective map for UCSD contains both a horizontal and vertical direction so we take the square root of the provided combined value. For training, we take 1600 random 79 × 119 pixel patches and for testing, we split each test image up into quadrants (which have dimension 79 × 119). There are two different ways to split the dataset into training and testing sets. We have experimented on the split that gave [\[18\]](#page-8-4) the best results as well as the split used in [\[28\]](#page-9-0).
 
 First, we split the dataset into four separate groups of training and testing sets as used in [\[18\]](#page-8-4) and originally defined by [\[20\]](#page-8-0). These groups are "upscale," "maximal," "minimal," and "downscale." We see in Table [3](#page-6-0) that the "upscale" split and "downscale" split give us state of the art results on counting for this dataset. For this experiment, we sampled 1600 random patches of size 119 × 79 pixels (width and height respectively) for the training set and split the test set images into 119 × 79 quadrants that could be reconstructed by piecing them together without overlap. We also added left-right flips of each image to our training data.
 
@@ -118,26 +118,26 @@ We then evaluate the original split. For this experiment, we similarly sampled 1
 
 The WorldExpo dataset [\[27\]](#page-8-7) contains a larger number of people (approximately 50 on average, which is double that of UCSD) and contains images from multiple locations. Perspective effects are also much more noticeable in this dataset as compared to UCSD. These qualities of the dataset serve to increase the difficulty of counting. Like UCSD, the WorldExpo dataset was constructed from frames of video recordings of crowds. This means that, unlike UCF, this dataset contains a relatively large number of training and testing images. We experiment on this dataset with and without perspective information.
 
-Without perspective maps, we generate label density maps for this dataset in the same manner as previously described: a 2D Gaussian with $\sigma$ = 15. We take 16000 150 × 150 randomly sampled patches for training. For testing, we densely scan the image, producing 150 × 150 patches at a stride of 100.
+Without perspective maps, we generate label density maps for this dataset in the same manner as previously described: a 2D Gaussian with $\sigma = 15$. We take 16000 150 × 150 randomly sampled patches for training. For testing, we densely scan the image, producing 150 × 150 patches at a stride of 100.
 
-When perspective maps are used, however, we follow the procedure as described in [\[27\]](#page-8-7), which involves estimating a "crowd density distribution kernel" as the sum of two 2D Gaussians: a symmetric Gaussian for the head and an ellipsoid Gaussian for the body. These are scaled by the perspective map $M$ provided, where $M(x)$ gives the number of pixels that represents a meter at pixel x [\[27\]](#page-8-7). Note that the meaning of this perspective map is distinct from the meaning of the perspective map provided for the UCSD dataset. Using this information, the density contribution from a person with head pixel x is given by the following sum of normalized Gaussians:
+When perspective maps are used, however, we follow the procedure as described in [\[27\]](#page-8-7), which involves estimating a “crowd density distribution kernel” as the sum of two 2D Gaussians: a symmetric Gaussian for the head and an ellipsoid Gaussian for the body. These are scaled by the perspective map $M$ provided, where $M(x)$ gives the number of pixels that represents a meter at pixel $x$ [\[27\]](#page-8-7). Note that the meaning of this perspective map is distinct from the meaning of the perspective map provided for the UCSD dataset. Using this information, the density contribution from a person with head pixel $x$ is given by the following sum of normalized Gaussians:
 
 $$D_{\mathbf{x}} = \frac{1}{||Z||} (\mathcal{N}_h(\mathbf{x}, \sigma_h) + \mathcal{N}_b(\mathbf{x}_b, \Sigma_b)) \qquad (5)$$
 
-where $x_b$ is the center of the body, which is 0.875 meters down from the head on average, and can be determined from the perspective map M and the head center x [\[27\]](#page-8-7). We sum these Gaussians for each person to pro-
+where $x_b$ is the center of the body, which is 0.875 meters down from the head on average, and can be determined from the perspective map $M$ and the head center $x$ [\[27\]](#page-8-7). We sum these Gaussians for each person to pro-
 
-| Method       | MAE           |
-|--------------|---------------|
-| AMDCN        | <b>290.82</b> |
-| Hydra2s [18] | 333.73        |
-| MCNN [28]    | 377.60        |
-| [27]         | 467.00        |
-| [23]         | 295.80        |
-| [3]          | 318.10        |
+| Method       | MAE    |
+|--------------|--------|
+| AMDCN        | 290.82 |
+| Hydra2s [18] | 333.73 |
+| MCNN [28]    | 377.60 |
+| [27]         | 467.00 |
+| [23]         | 295.80 |
+| [3]          | 318.10 |
 
 <span id="page-5-1"></span>Table 1. Mean absolute error of various methods on UCF crowds
 
-duce the final density map. We set $\sigma = 0.2M(x)$ for $N_h$ and $\sigma_x = 0.2M(x), \sigma_y = 0.5M(x)$ for $\Sigma_b$ in $N_b$.
+duce the final density map. We set $\sigma = 0.2M(\mathbf{x})$ for $N_h $and $\sigma_x = 0.2M(\mathbf{x}), \sigma_y = 0.5M(\mathbf{x})$ for $\Sigma_b$ in $N_b$.
 
 # 4. Results
 
@@ -149,19 +149,19 @@ We report a state of the art result on this dataset in Table [1,](#page-5-1) fol
 
 #### 4.2. TRANCOS Traffic Counting
 
-Our network performs very well on the TRANCOS dataset. Indeed, as confirmed by the GAME score, AMDCN produces the most accurate count and shape combined as compared to other methods. Table [2](#page-5-2) shows that we achieve state of the art results as measured by the [GAME](#GAME) metric [\[14\]](#page-8-1) across all levels.
+Our network performs very well on the TRANCOS dataset. Indeed, as confirmed by the GAME score, AMDCN produces the most accurate count and shape combined as compared to other methods. Table 2 shows that we achieve state of the art results as measured by the GAME metric [\[14\]](#page-8-1) across all levels.
 
 #### 4.3. UCSD Crowd Counting
 
 Results are shown in Table [3](#page-6-0) and Figure [3.](#page-6-1) We see that the "original" split as defined by the creators of the dataset in [\[5\]](#page-8-17) and used in [\[28\]](#page-9-0) gives us somewhat worse results for counting on this dataset. Results were consistent over multiple trainings. Again, including the perspective map does not seem to increase performance on this dataset. Despite this, we see in Table [3](#page-6-0) and Figure [3](#page-6-1) that the results are comparable to the state of the art. In fact, for two of the splits, our proposed network beats the state of the art. For the upscale split, the AMDCN is the state of the art by a large relative margin. This is compelling because it shows that accurate perspective-free counting can be achieved without
 
-| Method                   | GAME<br>(L=0)                             | GAME<br>(L=1) | GAME<br>(L=2) | GAME<br>(L=3) |       |
-|--------------------------|-------------------------------------------|---------------|---------------|---------------|-------|
-| AMDCN                    | 9.77                                      | 13.16         | 15.00         | 15.87         |       |
-| [18]                     | 10.99                                     | 13.75         | 16.69         | 19.32         |       |
-| [15] + SIFT<br>from [14] | 13.76                                     | 16.72         | 20.72         | 24.36         |       |
-|                          | [13] + RGB<br>Norm + Filters<br>from [14] | 17.68         | 19.97         | 23.54         | 25.84 |
-| HOG-2<br>from [14]       |                                           | 13.29         | 18.05         | 23.65         | 28.41 |
+| Method                                    | GAME<br>(L=0) | GAME<br>(L=1) | GAME<br>(L=2) | GAME<br>(L=3) |
+|-------------------------------------------|---------------|---------------|---------------|---------------|
+| AMDCN                                     | <b>9.77</b>   | <b>13.16</b>  | <b>15.00</b>  | <b>15.87</b>  |
+| [18]                                      | 10.99         | 13.75         | 16.69         | 19.32         |
+| [15] + SIFT<br>from [14]                  | 13.76         | 16.72         | 20.72         | 24.36         |
+| [13] + RGB<br>Norm + Filters<br>from [14] | 17.68         | 19.97         | 23.54         | 25.84         |
+| HOG-2<br>from [14]                        | 13.29         | 18.05         | 23.65         | 28.41         |
 
 <span id="page-5-2"></span>Table 2. Mean absolute error of various methods on TRANCOS traffic
 
@@ -181,22 +181,22 @@ We report the results of the ablation studies in Figure [4.](#page-7-2) We note
 
 <span id="page-6-1"></span>Figure 3. UCSD crowd counting dataset. Both plots show comparisons of predicted and ground truth counts over time. While AMDCN does not beat the state of the art on the original split, the predictions still follow the true counts reasonably. The jump in the original split is due to that testing set including multiple scenes of highly varying counts.
 
-| Method                                  | maximal     | downscale   | upscale     | minimal     | original    |
-|-----------------------------------------|-------------|-------------|-------------|-------------|-------------|
-| AMDCN (without perspective information) | 1.63        | 1.43        | <b>0.63</b> | 1.71        | 1.74        |
-| AMDCN (with perspective information)    | 1.60        | <b>1.24</b> | 1.37        | 1.59        | 1.72        |
-| [18] (with perspective information)     | 1.65        | 1.79        | 1.11        | 1.50        | -           |
-| [18] (without perspective information)  | 2.22        | 1.93        | 1.37        | 2.38        | -           |
-| [15]                                    | 1.70        | 1.28        | 1.59        | 2.02        | -           |
-| [13]                                    | 1.70        | 2.16        | 1.61        | 2.20        | -           |
-| [19]                                    | 1.43        | 1.30        | 1.59        | 1.62        | -           |
-| [2]                                     | <b>1.24</b> | 1.31        | 1.69        | <b>1.49</b> | -           |
-| [27]                                    | 1.70        | 1.26        | 1.59        | 1.52        | 1.60        |
-| [28]                                    | -           | -           | -           | -           | <b>1.07</b> |
-| [1, 28]                                 | -           | -           | -           | -           | 2.16        |
-| [7]                                     | -           | -           | -           | -           | 2.25        |
-| [5]                                     | -           | -           | -           | -           | 2.24        |
-| [6]                                     | -           | -           | -           | -           | 2.07        |
+| Method                                  | maximal | downscale | upscale | minimal | original |
+|-----------------------------------------|---------|-----------|---------|---------|----------|
+| AMDCN (without perspective information) | 1.63    | 1.43      | 0.63    | 1.71    | 1.74     |
+| AMDCN (with perspective information)    | 1.60    | 1.24      | 1.37    | 1.59    | 1.72     |
+| [18] (with perspective information)     | 1.65    | 1.79      | 1.11    | 1.50    | -        |
+| [18] (without perspective information)  | 2.22    | 1.93      | 1.37    | 2.38    | -        |
+| [15]                                    | 1.70    | 1.28      | 1.59    | 2.02    | -        |
+| [13]                                    | 1.70    | 2.16      | 1.61    | 2.20    | -        |
+| [19]                                    | 1.43    | 1.30      | 1.59    | 1.62    | -        |
+| [2]                                     | 1.24    | 1.31      | 1.69    | 1.49    | -        |
+| [27]                                    | 1.70    | 1.26      | 1.59    | 1.52    | 1.60     |
+| [28]                                    | -       | -         | -       | -       | 1.07     |
+| [1, 28]                                 | -       | -         | -       | -       | 2.16     |
+| [7]                                     | -       | -         | -       | -       | 2.25     |
+| [5]                                     | -       | -         | -       | -       | 2.24     |
+| [6]                                     | -       | -         | -       | -       | 2.07     |
 
 <span id="page-6-0"></span>Table 3. Mean absolute error of various methods on UCSD crowds
 
@@ -267,4 +267,4 @@ counting. In *Proceedings of the IEEE Conference on Computer Vision and Pattern
 
 *Computer Vision and Pattern Recognition*, pages 833– 841, 2015.
 
-- <span id="page-9-0"></span>[28] Y. Zhang, D. Zhou, S. Chen, S. Gao, and Y. Ma. Single-image crowd counting via multi-column convolutional neural network. In *Proceedings of the IEEE Conference on Computer Vision and Pattern Recogni tion*, pages 589–597, 2016.
\ No newline at end of file
+- <span id="page-9-0"></span>[\[28\]](#page-8-8) Y. Zhang, D. Zhou, S. Chen, S. Gao, and Y. Ma. Single-image crowd counting via multi-column convolutional neural network. In *Proceedings of the IEEE Conference on Computer Vision and Pattern Recogni tion*, pages 589–597, 2016.
\ No newline at end of file
diff --git a/data/examples/markdown/switch_transformers/switch_trans.md b/data/examples/markdown/switch_transformers/switch_trans.md
index fcec2feb..270b99cf 100644
--- a/data/examples/markdown/switch_transformers/switch_trans.md
+++ b/data/examples/markdown/switch_transformers/switch_trans.md
@@ -16,7 +16,7 @@ Editor: Alexander Clark
 
 ### Abstract
 
-In deep learning, models typically reuse the same parameters for all inputs. Mixture of Experts (MoE) models defy this and instead select different parameters for each incoming example. The result is a sparsely-activated model—with an outrageous number of parameters—but a constant computational cost. However, despite several notable successes of MoE, widespread adoption has been hindered by complexity, communication costs, and training instability. We address these with the introduction of the Switch Transformer. We simplify the MoE routing algorithm and design intuitive improved models with reduced communication and computational costs. Our proposed training techniques mitigate the instabilities, and we show large sparse models may be trained, for the first time, with lower precision (bfloat16) formats. We design models based off T5-Base and T5-Large [(Raffel](#page-37-0) [et al.,](#page-37-0) [2019)](#page-37-0) to obtain up to 7x increases in pre-training speed with the same computational resources. These improvements extend into multilingual settings where we measure gains over the mT5-Base version across all 101 languages. Finally, we advance the current scale of language models by pre-training up to trillion parameter models on the "Colossal Clean Crawled Corpus", and achieve a 4x speedup over the T5-XXL model.[1](#page-0-0)[2](#page-0-1)
+In deep learning, models typically reuse the same parameters for all inputs. Mixture of Experts (MoE) models defy this and instead select different parameters for each incoming example. The result is a sparsely-activated model—with an outrageous number of parameters—but a constant computational cost. However, despite several notable successes of MoE, widespread adoption has been hindered by complexity, communication costs, and training instability. We address these with the introduction of the Switch Transformer. We simplify the MoE routing algorithm and design intuitive improved models with reduced communication and computational costs. Our proposed training techniques mitigate the instabilities, and we show large sparse models may be trained, for the first time, with lower precision (bfloat16) formats. We design models based off T5-Base and T5-Large (Raffel et al., 2019) to obtain up to 7x increases in pre-training speed with the same computational resources. These improvements extend into multilingual settings where we measure gains over the mT5-Base version across all 101 languages. Finally, we advance the current scale of language models by pre-training up to trillion parameter models on the “Colossal Clean Crawled Corpus”, and achieve a 4x speedup over the T5-XXL model.[1](#page-0-0)[2](#page-0-1)
 
 Keywords: mixture-of-experts, natural language processing, sparsity, large-scale machine learning, distributed computing
 
@@ -30,22 +30,38 @@ Keywords: mixture-of-experts, natural language processing, sparsity, large-scale
 
 ## Contents
 
-| 1 | Introduction                                                                                                                                                                                                                                                       | 3                                      |
-|---|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|----------------------------------------|
-| 2 | Switch Transformer<br>2.1 Simplifying Sparse Routing<br>2.2 Efficient Sparse Routing<br>2.3 Putting It All Together: The Switch Transformer<br>2.4 Improved Training and Fine-Tuning Techniques                                                                    | 4<br>5<br>6<br>8<br>8                  |
-| 3 | Scaling Properties<br>3.1 Scaling Results on a Step-Basis<br>3.2 Scaling Results on a Time-Basis<br>3.3 Scaling Versus a Larger Dense Model                                                                                                                        | 11<br>12<br>13<br>13                   |
-| 4 | Downstream Results<br>4.1 Fine-Tuning<br>4.2 Distillation<br>4.3 Multilingual Learning                                                                                                                                                                             | 14<br>14<br>16<br>17                   |
-| 5 | Designing Models with Data, Model, and Expert-Parallelism<br>5.1 Data Parallelism<br>5.2 Model Parallelism<br>5.3 Model and Data Parallelism<br>5.4 Expert and Data Parallelism<br>5.5 Expert, Model and Data Parallelism<br>5.6 Towards Trillion Parameter Models | 18<br>20<br>20<br>21<br>22<br>22<br>22 |
-| 6 | Related Work                                                                                                                                                                                                                                                       | 24                                     |
-| 7 | Discussion                                                                                                                                                                                                                                                         | 25                                     |
-| 8 | Future Work                                                                                                                                                                                                                                                        | 26                                     |
-| 9 | Conclusion                                                                                                                                                                                                                                                         | 27                                     |
-| A | Switch for Attention                                                                                                                                                                                                                                               | 27                                     |
-| B | Preventing Token Dropping with <i>No-Token-Left-Behind</i>                                                                                                                                                                                                         | 29                                     |
-| C | Encouraging Exploration Across Experts                                                                                                                                                                                                                             | 29                                     |
-| D | Switch Transformers in Lower Compute Regimes                                                                                                                                                                                                                       | 29                                     |
-| E | Relation of Upstream to Downstream Model Performance                                                                                                                                                                                                               | 32                                     |
-| F | Pseudo Code for Switch Transformers                                                                                                                                                                                                                                | 33                                     |
+| 1 | Introduction                                              | 3  |    |
+|---|-----------------------------------------------------------|----|----|
+| 2 | Switch Transformer                                        | 4  |    |
+|   | 2.1 Simplifying Sparse Routing                            | 5  |    |
+|   | 2.2 Efficient Sparse Routing                              | 6  |    |
+|   | 2.3 Putting It All Together: The Switch Transformer       | 8  |    |
+|   | 2.4 Improved Training and Fine-Tuning Techniques          | 8  |    |
+| 3 | Scaling Properties                                        | 11 |    |
+|   | 3.1 Scaling Results on a Step-Basis                       | 12 |    |
+|   | 3.2 Scaling Results on a Time-Basis                       | 13 |    |
+|   | 3.3 Scaling Versus a Larger Dense Model                   | 13 |    |
+| 4 | Downstream Results                                        | 14 |    |
+|   | 4.1 Fine-Tuning                                           | 14 |    |
+|   | 4.2 Distillation                                          | 16 |    |
+|   | 4.3 Multilingual Learning                                 | 17 |    |
+| 5 | Designing Models with Data, Model, and Expert-Parallelism | 18 |    |
+|   | 5.1 Data Parallelism                                      | 20 |    |
+|   | 5.2 Model Parallelism                                     | 20 |    |
+|   | 5.3 Model and Data Parallelism                            | 21 |    |
+|   | 5.4 Expert and Data Parallelism                           | 22 |    |
+|   | 5.5 Expert, Model and Data Parallelism                    | 22 |    |
+|   | 5.6 Towards Trillion Parameter Models                     | 22 |    |
+| 6 | Related Work                                              | 24 |    |
+| 7 | Discussion                                                | 25 |    |
+| 8 | Future Work                                               | 26 |    |
+| 9 | Conclusion                                                | 27 |    |
+| A | Switch for Attention                                      | 27 |    |
+| B | Preventing Token Dropping with No-Token-Left-Behind       | 29 |    |
+| C | Encouraging Exploration Across Experts                    | 29 |    |
+| D | Switch Transformers in Lower Compute Regimes              |    | 29 |
+| E | Relation of Upstream to Downstream Model Performance      |    | 32 |
+| F | Pseudo Code for Switch Transformers                       | 33 |    |
 
 ### <span id="page-2-0"></span>1. Introduction
 
@@ -74,24 +90,26 @@ Heeding these results, we investigate a fourth axis: increase the parameter coun
 
 ![](_page_4_Figure_1.jpeg)
 
+Figure 2: Illustration of a Switch Transformer encoder block. We replace the dense feed forward network (FFN) layer present in the Transformer with a sparse Switch FFN layer (light blue). The layer operates independently on the tokens in the sequence. We diagram two tokens ($x_1$ = “More” and $x_2$ = “Parameters” below) being routed (solid lines) across four FFN experts, where the router independently routes each token. The switch FFN layer returns the output of the selected FFN multiplied by the router gate value (dotted-line).
+
 ### <span id="page-4-0"></span>2.1 Simplifying Sparse Routing
 
-Mixture of Expert Routing. [Shazeer et al.](#page-38-2) [(2017)](#page-38-2) proposed a natural language Mixtureof-Experts (MoE) layer which takes as an input a token representation x and then routes this to the best determined top-$k$ experts, selected from a set ${E_i(x)}_{i=1}^N$ of $N$ experts. The router variable $W_r$ produces logits $h(x) = W_r \cdot x$ which are normalized via a softmax distribution over the available N experts at that layer. The gate-value for expert i is given by,
+Mixture of Expert Routing. [Shazeer et al.](#page-38-2) [(2017)](#page-38-2) proposed a natural language Mixtureof-Experts (MoE) layer which takes as an input a token representation $x$ and then routes this to the best determined top-$k$ experts, selected from a set ${E_i(x)\}_{i=1}^N$ of $N$ experts. The router variable $W_r$ produces logits $h(x) = W_r \cdot x$ which are normalized via a softmax distribution over the available $N$ experts at that layer. The gate-value for expert $i$ is given by,
 
 $$p_i(x) = \frac{e^{h(x)_i}}{\sum_j^N e^{h(x)_j}}.\tag{1}$$
 
-The top-$k$ gate values are selected for routing the token $x$. If $\mathcal{T}$ is the set of selected top-$k $indices then the output computation of the layer is the linearly weighted combination of each expert's computation on the token by the gate value,
+The top-$k$ gate values are selected for routing the token $x$. If $\mathcal{T}$ is the set of selected top-$k$ indices then the output computation of the layer is the linearly weighted combination of each expert's computation on the token by the gate value,
 
 <span id="page-4-1"></span>
 $$y = \sum_{i \in \mathcal{T}} p_i(x) E_i(x). \tag{2}$$
 
-Switch Routing: Rethinking Mixture-of-Experts. [Shazeer et al.](#page-38-2) [(2017)](#page-38-2) conjectured that routing to $k > 1$ experts was necessary in order to have non-trivial gradients to the routing functions. The authors intuited that learning to route would not work without the ability to compare at least two experts. [Ramachandran and Le](#page-37-3) [(2018)](#page-37-3) went further to study the top-k decision and found that higher k-values in lower layers in the model were important for models with many routing layers. Contrary to these ideas, we instead use a simplified strategy where we route to only a single expert. We show this simplification preserves model quality, reduces routing computation and performs better. This $k = 1 $routing strategy is later referred to as a Switch layer. Note that for both MoE and Switch Routing, the gate value $p_i(x)$ in Equation [2](#page-4-1) permits differentiability of the router.
+**Switch Routing:** Rethinking Mixture-of-Experts. [Shazeer et al.](#page-38-2) (2017) conjectured that routing to $k > 1$ experts was necessary in order to have non-trivial gradients to the routing functions. The authors intuited that learning to route would not work without the ability to compare at least two experts. [Ramachandran and Le](#page-37-3) (2018) went further to study the top-$k$ decision and found that higher $k$-values in lower layers in the model were important for models with many routing layers. Contrary to these ideas, we instead use a simplified strategy where we route to only a single expert. We show this simplification preserves model quality, reduces routing computation and performs better. This $k = 1 $routing strategy is later referred to as a Switch layer. Note that for both MoE and Switch Routing, the gate value $p_i(x)$ in Equation [2](#page-4-1) permits differentiability of the router.
 
 The benefits for the Switch layer are three-fold: (1) The router computation is reduced as we are only routing a token to a single expert. (2) The batch size (expert capacity) of each expert can be at least halved since each token is only being routed to a single expert.[3](#page-5-1) (3) The routing implementation is simplified and communication costs are reduced. Figure [3](#page-5-2) shows an example of routing with different expert capacity factors.
 
 ![](_page_5_Figure_3.jpeg)
 
-- <span id="page-5-2"></span>Figure 3: Illustration of token routing dynamics. Each expert processes a fixed batch-size of tokens modulated by the capacity factor. Each token is routed to the expert with the highest router probability, but each expert has a fixed batch size of (total_tokens / num_experts) $\times$ capacity_factor. If the tokens are unevenly dispatched then certain experts will overflow (denoted by dotted red lines), resulting in these tokens not being processed by this layer. A larger capacity factor alleviates this overflow issue, but also increases computation and communication costs (depicted by padded white/empty slots).
+- <span id="page-5-2"></span>Figure 3: Illustration of token routing dynamics. Each expert processes a fixed batch-size of tokens modulated by the capacity factor. Each token is routed to the expert with the highest router probability, but each expert has a fixed batch size of (total_tokens / num_experts) × capacity_factor. If the tokens are unevenly dispatched then certain experts will overflow (denoted by dotted red lines), resulting in these tokens not being processed by this layer. A larger capacity factor alleviates this overflow issue, but also increases computation and communication costs (depicted by padded white/empty slots).
 ### <span id="page-5-0"></span>2.2 Efficient Sparse Routing
 
 We use Mesh-Tensorflow (MTF) [(Shazeer et al.,](#page-38-3) [2018)](#page-38-3) which is a library, with similar semantics and API to Tensorflow [(Abadi et al.,](#page-35-3) [2016)](#page-35-3) that facilitates efficient distributed data and model parallel architectures. It does so by abstracting the physical set of cores to a logical mesh of processors. Tensors and computations may then be sharded per named dimensions, facilitating easy partitioning of models across dimensions. We design our model with TPUs in mind, which require statically declared sizes. Below we describe our distributed Switch Transformer implementation.
@@ -102,25 +120,25 @@ Distributed Switch Implementation. All of our tensor shapes are statically deter
 
 $$\text{expected capacity} = \left(\frac{\text{tokens per batch}}{\text{number of experts}}\right) \times \text{capacity factor}.\tag{3}$$
 
-A capacity factor greater than 1.0 creates additional buffer to accommodate for when tokens are not perfectly balanced across experts. If too many tokens are routed to an expert (referred to later as dropped tokens), computation is skipped and the token representation is passed directly to the next layer through the residual connection. Increasing the expert capacity is not without drawbacks, however, since high values will result in wasted computation and memory. This trade-off is explained in Figure [3.](#page-5-2) Empirically we find ensuring lower rates of dropped tokens are important for the scaling of sparse expert-models. Throughout our experiments we didn't notice any dependency on the number of experts for the number of tokens dropped (typically < 1%). Using the auxiliary load balancing loss (next section) with a high enough coefficient ensured good load balancing. We study the impact that these design decisions have on model quality and speed in Table [1.](#page-8-0)
+A capacity factor greater than 1.0 creates additional buffer to accommodate for when tokens are not perfectly balanced across experts. If too many tokens are routed to an expert (referred to later as dropped tokens), computation is skipped and the token representation is passed directly to the next layer through the residual connection. Increasing the expert capacity is not without drawbacks, however, since high values will result in wasted computation and memory. This trade-off is explained in Figure [3.](#page-5-2) Empirically we find ensuring lower rates of dropped tokens are important for the scaling of sparse expert-models. Throughout our experiments we didn't notice any dependency on the number of experts for the number of tokens dropped (typically $< 1%$). Using the auxiliary load balancing loss (next section) with a high enough coefficient ensured good load balancing. We study the impact that these design decisions have on model quality and speed in Table [1.](#page-8-0) 
 
-A Differentiable Load Balancing Loss. To encourage a balanced load across experts we add an auxiliary loss [(Shazeer et al.,](#page-38-2) [2017,](#page-38-2) [2018;](#page-38-3) [Lepikhin et al.,](#page-37-2) [2020)](#page-37-2). As in [Shazeer](#page-38-3) [et al.](#page-38-3) [(2018)](#page-38-3); [Lepikhin et al.](#page-37-2) [(2020)](#page-37-2), Switch Transformers simplifies the original design in [Shazeer et al.](#page-38-2) [(2017)](#page-38-2) which had separate load-balancing and importance-weighting losses. For each Switch layer, this auxiliary loss is added to the total model loss during training. Given $N$ experts indexed by $i = 1$ to $N$ and a batch $\mathcal{B}$ with $T$ tokens, the auxiliary loss is computed as the scaled dot-product between vectors $f$ and $P$,
+A Differentiable Load Balancing Loss. To encourage a balanced load across experts we add an auxiliary loss (Shazeer et al., 2017, 2018; Lepikhin et al., 2020). As in Shazeer et al. (2018); Lepikhin et al. (2020), Switch Transformers simplifies the original design in Shazeer et al. (2017) which had separate load-balancing and importance-weighting losses. For each Switch layer, this auxiliary loss is added to the total model loss during training. Given $N$ experts indexed by $i = 1$ to $N$ and a batch $\mathcal{B}$ with $T$ tokens, the auxiliary loss is computed as the scaled dot-product between vectors $f$ and $P$,
 
 $$\text{loss} = \alpha \cdot N \cdot \sum_{i=1}^{N} f_i \cdot P_i \tag{4}$$
 
-<span id="page-6-1"></span>where fi is the fraction of tokens dispatched to expert i,
+<span id="page-6-1"></span>where $f_i$ is the fraction of tokens dispatched to expert $i$,
 
 $$f_i = \frac{1}{T} \sum_{x \in \mathcal{B}} \mathbb{1} \{ \operatorname*{argmax} \, p(x) = i \} \tag{5}$$
 
-and Pi is the fraction of the router probability allocated for expert i, [2](#page-6-0)
+and $P_i$ is the fraction of the router probability allocated for expert $i$,2
 
 $$P_i = \frac{1}{T} \sum_{x \in \mathcal{B}} p_i(x). \tag{6}$$
 
-Since we seek uniform routing of the batch of tokens across the N experts, we desire both vectors to have values of $1/N$. The auxiliary loss of Equation [4](#page-6-1) encourages uniform routing since it is minimized under a uniform distribution. The objective can also be differentiated as
+Since we seek uniform routing of the batch of tokens across the $N$ experts, we desire both vectors to have values of $1/N$. The auxiliary loss of Equation [4](#page-6-1) encourages uniform routing since it is minimized under a uniform distribution. The objective can also be differentiated as
 
-<span id="page-6-0"></span>2. A potential source of confusion: $p_i(x)$ is the probability of routing token $x$ to expert $i$. $P_i$ is the probability fraction to expert i across all tokens in the batch B.
+<span id="page-6-0"></span>A potential source of confusion: $p_i(x)$ is the probability of routing token $x$ to expert $i$. $P_i$ is the probability fraction to expert $i$ across all tokens in the batch $B$.
 
-the P-vector is differentiable, but the f-vector is not. The final loss is multiplied by expert countNto keep the loss constant as the number of experts varies since under uniform$\sum_{i=1}^{N} p_i(x) - \sum_{i=1}^{N} \frac{N}{N}$. Many, a hyper-parameter $α$ is a multiplicative · sufficiently large to ensure load balancing while small enough to not to overwhelm the coefficient for these auxiliary losses; throughout this work we use an $α = 10^{-2}$ which was primary cross-entropy objective. We swept hyper-parameter ranges of $α$ from $10^{-1}$ to $10^{-5}$
+the $P$-vector is differentiable, but the $f$-vector is not. The final loss is multiplied by expert count $N$ to keep the loss constant as the number of experts varies since under uniform routing $\sum_{i=1}^{N}(f_i$ · Pi) = PN i=1( 1 N · 1 N ) = 1 N . Finally, a hyper-parameter α is a multiplicative coefficient for these auxiliary losses; throughout this work we use an α = 10−2 which was sufficiently large to ensure load balancing while small enough to not to overwhelm the primary cross-entropy objective. We swept hyper-parameter ranges of $\alpha$ from $10^{-1}$ to $10^{-5}$ in powers of 10 and found $10^{-2}$ balanced load quickly without interfering with training loss.
 
 #### <span id="page-7-0"></span>2.3 Putting It All Together: The Switch Transformer
 
@@ -157,20 +175,19 @@ Selective precision with large sparse models. Model instability hinders the abil
 
 To achieve this, we cast the router input to float32 precision. The router function takes the tokens as input and produces the dispatch and combine tensors used for the selection and recombination of expert computation (refer to Code Block [15](#page-33-0) in the Appendix for details). Importantly, the float32 precision is only used within the body of the router function—on computations local to that device. Because the resulting dispatch and combine tensors are recast to bfloat16 precision at the end of the function, no expensive float32 tensors
 
-| Model                             | Quality              | Speed              |
-|-----------------------------------|----------------------|--------------------|
-| (precision)                       | (Neg. Log Perp.) (↑) | (Examples/sec) (↑) |
-| Switch-Base (float32)             | -1.718               | 1160               |
-| Switch-Base (bfloat16)            | -3.780 [diverged]    | <b>1390</b>        |
-| Switch-Base (Selective precision) | <b>-1.716</b>        | 1390               |
+| Model<br>(precision)              | Quality<br>(Neg. Log Perp.) (↑) | Speed<br>(Examples/sec) (↑) |
+|-----------------------------------|---------------------------------|-----------------------------|
+| Switch-Base (float32)             | -1.718                          | 1160                        |
+| Switch-Base (bfloat16)            | -3.780 [diverged]               | <b>1390</b>                 |
+| Switch-Base (Selective precision) | <b>-1.716</b>                   | 1390                        |
 
 <span id="page-9-0"></span>Table 2: Selective precision. We cast the local routing operations to float32 while preserving bfloat16 precision elsewhere to stabilize our model while achieving nearly equal speed to (unstable) bfloat16-precision training. We measure the quality of a 32 expert model after a fixed step count early in training its speed performance. For both Switch-Base in float32 and with Selective prevision we notice similar learning dynamics.
 
 are broadcast through all-to-all communication operations, but we still benefit from the increased stability of float32.
 
-Smaller parameter initialization for stability. Appropriate initialization is critical to successful training in deep learning and we especially observe this to be true for Switch Transformer. We initialize our weight matrices by drawing elements from a truncated in powers of 10 and found $10^{-2}$ balanced load quickly without interfering with training loss. hyper-parameter and n is the number of input units in the weight tensor (e.g. fan-in).[6](#page-9-1)
+Smaller parameter initialization for stability. Appropriate initialization is critical to successful training in deep learning and we especially observe this to be true for Switch Transformer. We initialize our weight matrices by drawing elements from a truncated normal distribution with mean $\mu = 0$ and standard deviation $\sigma = \sqrt{s/n}$ where $s$ is a scale hyper-parameter and $n$ is the number of input units in the weight tensor (e.g. fan-in).[6](#page-9-1)
 
-As an additional remedy to the instability, we recommend reducing the default Transnormal distribution with mean $μ = 0$ and standard deviation $σ = \sqrt{s/n}$ where $s$ is a scale the likelihood of destabilized training in our experiments. Table [3](#page-9-2) measures the improvement of the model quality and reduction of the variance early in training. We find that
+As an additional remedy to the instability, we recommend reducing the default Transformer initialization scale $s = 1.0$ by a factor of 10. This both improves quality and reduces the likelihood of destabilized training in our experiments. Table [3](#page-9-2) measures the improvement of the model quality and reduction of the variance early in training. We find that
 
 | Model (Initialization scale) | Average Quality<br>(Neg. Log Perp.) | Std. Dev. of Quality<br>(Neg. Log Perp.) |
 |------------------------------|-------------------------------------|------------------------------------------|
@@ -184,18 +201,17 @@ the average model quality, as measured by the Neg. Log Perp., is dramatically im
 
 Regularizing large sparse models. Our paper considers the common NLP approach of pre-training on a large corpus followed by fine-tuning on smaller downstream tasks such as summarization or question answering. One issue that naturally arises is overfitting since many fine-tuning tasks have very few examples. During fine-tuning of standard Transformers, [Raffel et al.](#page-37-0) [(2019)](#page-37-0) use dropout [(Srivastava et al.,](#page-38-5) [2014)](#page-38-5) at each layer to prevent overfitting. Our Switch Transformers have significantly more parameters than the FLOP matched dense baseline, which can lead to more severe overfitting on these smaller downstream tasks.
 
-| Model (dropout)             | GLUE | CNNDM | SQuAD | SuperGLUE |
-|-----------------------------|------|-------|-------|-----------|
-| T5-Base (d=0.1)             | 82.9 | 19.6  | 83.5  | 72.4      |
-| Switch-Base (d=0.1)         | 84.7 | 19.1  | 83.7  | 73.0      |
-| Switch-Base (d=0.2)         | 84.4 | 19.2  | 83.9  | 73.2      |
-| Switch-Base (d=0.3)         | 83.9 | 19.6  | 83.4  | 70.7      |
-| Switch-Base (d=0.1, ed=0.4) | 85.2 | 19.6  | 83.7  | 73.0      |
+| Model (dropout)             | GLUE        | CNNDM       | SQuAD       | SuperGLUE   |
+|-----------------------------|-------------|-------------|-------------|-------------|
+| T5-Base (d=0.1)             | 82.9        | <b>19.6</b> | 83.5        | 72.4        |
+| Switch-Base (d=0.1)         | 84.7        | 19.1        | <b>83.7</b> | <b>73.0</b> |
+| Switch-Base (d=0.2)         | 84.4        | 19.2        | <b>83.9</b> | <b>73.2</b> |
+| Switch-Base (d=0.3)         | 83.9        | 19.6        | 83.4        | 70.7        |
+| Switch-Base (d=0.1, ed=0.4) | <b>85.2</b> | <b>19.6</b> | <b>83.7</b> | <b>73.0</b> |
 
 <span id="page-10-1"></span>Table 4: Fine-tuning regularization results. A sweep of dropout rates while fine-tuning Switch Transformer models pre-trained on 34B tokens of the C4 data set (higher numbers are better). We observe that using a lower standard dropout rate at all non-expert layer, with a much larger dropout rate on the expert feed-forward layers, to perform the best.
 
-We thus propose a simple way to alleviate this issue during fine-tuning: increase the dropout inside the experts, which we name as expert dropout. During fine-tuning we simply increase the dropout rate by a significant amount only at the interim feed-forward computation at each expert layer. Table [4](#page-10-1) has the results for our expert dropout protocol. We observe that simply increasing the dropout across all layers leads to worse performance. former initialization scale $s = 1.0$ by a factor of 10. This both improves quality and reduces However, setting a smaller dropout rate (0.1) at non-expert layers and a much larger dropout
-rate (0.4) at expert layers leads to performance improvements on four smaller downstream tasks.
+We thus propose a simple way to alleviate this issue during fine-tuning: increase the dropout inside the experts, which we name as expert dropout. During fine-tuning we simply increase the dropout rate by a significant amount only at the interim feed-forward computation at each expert layer. Table [4](#page-10-1) has the results for our expert dropout protocol. We observe that simply increasing the dropout across all layers leads to worse performance. However, setting a smaller dropout rate (0.1) at non-expert layers and a much larger dropout rate (0.4) at expert layers leads to performance improvements on four smaller downstream tasks.
 
 ### <span id="page-10-0"></span>3. Scaling Properties
 
@@ -236,11 +252,11 @@ Section [3](#page-10-0) demonstrated the superior scaling properties while pre-t
 
 ### <span id="page-13-1"></span>4.1 Fine-Tuning
 
-Baseline and Switch models used for fine-tuning. Our baselines are the highly-tuned 223M parameter T5-Base model and the 739M parameter T5-Large model [(Raffel et al.,](#page-37-0) [2019)](#page-37-0). For both versions, we design a FLOP-matched Switch Transformer, with many more parameters, which is summarized in Table [9.](#page-22-0) [7](#page-13-3) Our baselines differ slightly from those in [Raffel et al.](#page-37-0) [(2019)](#page-37-0) because we pre-train on an improved C4 corpus which removes intraexample text duplication and thus increases the efficacy as a pre-training task [Lee et al.](#page-37-5)
+Baseline and Switch models used for fine-tuning. Our baselines are the highly-tuned 223M parameter T5-Base model and the 739M parameter T5-Large model [(Raffel et al.,](#page-37-0) [2019)](#page-37-0). For both versions, we design a FLOP-matched Switch Transformer, with many more parameters, which is summarized in Table 9.[7](#page-13-3) Our baselines differ slightly from those in [Raffel et al.](#page-37-0) [(2019)](#page-37-0) because we pre-train on an improved C4 corpus which removes intraexample text duplication and thus increases the efficacy as a pre-training task [Lee et al.](#page-37-5)
 
 <span id="page-13-3"></span><sup>7.</sup> FLOPS are calculated for the forward pass as done in [Kaplan et al.](#page-36-0) [(2020)](#page-36-0).
 
-[(2021)](#page-37-5). In our protocol we pre-train with 220 (1,048,576) tokens per batch for 550k steps amounting to 576B total tokens. We then fine-tune across a diverse set of tasks using a dropout rate of 0.1 for all layers except the Switch layers, which use a dropout rate of 0.4 (see Table [4)](#page-10-1). We fine-tune using a batch-size of 1M for 16k steps and for each task, we evaluate model quality every 200-steps and report the peak performance as computed on the validation set.
+[(2021)](#page-37-5). In our protocol we pre-train with $2^{20}$ (1,048,576) tokens per batch for 550k steps amounting to 576B total tokens. We then fine-tune across a diverse set of tasks using a dropout rate of 0.1 for all layers except the Switch layers, which use a dropout rate of 0.4 (see Table [4](#page-10-1)). We fine-tune using a batch-size of 1M for 16k steps and for each task, we evaluate model quality every 200-steps and report the peak performance as computed on the validation set.
 
 Fine-tuning tasks and data sets. We select tasks probing language capabilities including question answering, summarization and knowledge about the world. The language benchmarks GLUE [(Wang et al.,](#page-39-4) [2018)](#page-39-4) and SuperGLUE [(Wang et al.,](#page-39-5) [2019)](#page-39-5) are handled as composite mixtures with all the tasks blended in proportion to the amount of tokens present in each. These benchmarks consist of tasks requiring sentiment analysis (SST-2), word sense disambiguation (WIC), sentence similarty (MRPC, STS-B, QQP), natural language inference (MNLI, QNLI, RTE, CB), question answering (MultiRC, RECORD, BoolQ), coreference resolution (WNLI, WSC) and sentence completion (COPA) and sentence acceptability (CoLA). The CNNDM [(Hermann et al.,](#page-36-4) [2015)](#page-36-4) and BBC XSum [(Narayan](#page-37-6) [et al.,](#page-37-6) [2018)](#page-37-6) data sets are used to measure the ability to summarize articles. Question answering is probed with the SQuAD data set [(Rajpurkar et al.,](#page-37-7) [2016)](#page-37-7) and the ARC Reasoning Challenge [(Clark et al.,](#page-35-6) [2018)](#page-35-6). And as in [Roberts et al.](#page-38-6) [(2020)](#page-38-6), we evaluate the knowledge of our models by fine-tuning on three closed-book question answering data sets: Natural Questions [(Kwiatkowski et al.,](#page-36-5) [2019)](#page-36-5), Web Questions [(Berant et al.,](#page-35-7) [2013)](#page-35-7) and Trivia QA [(Joshi et al.,](#page-36-6) [2017)](#page-36-6). Closed-book refers to questions posed with no supplemental reference or context material. To gauge the model's common sense reasoning we evaluate it on the Winogrande Schema Challenge [(Sakaguchi et al.,](#page-38-7) [2020)](#page-38-7). And finally, we test our model's natural language inference capabilities on the Adversarial NLI Benchmark [(Nie et al.,](#page-37-8) [2019)](#page-37-8).
 
@@ -248,7 +264,7 @@ Fine-tuning metrics. The following evaluation metrics are used throughout the pa
 
 Fine-tuning results. We observe significant downstream improvements across many natural language tasks. Notable improvements come from SuperGLUE, where we find FLOP-matched Switch variants improve by 4.4 and 2 percentage points over the T5-Base and T5-Large baselines, respectively as well as large improvements in Winogrande, closed book Trivia QA, and XSum.[8](#page-14-0) In our fine-tuning study, the only tasks where we do not observe gains are on the AI2 Reasoning Challenge (ARC) data sets where the T5-Base outperforms Switch-Base on the challenge data set and T5-Large outperforms Switch-Large on the easy data set. Taken as a whole, we observe significant improvements spanning both reasoning and knowledge-heavy tasks. This validates our architecture, not just as one that pre-trains well, but can translate quality improvements to downstream tasks via fine-tuning.
 
-<span id="page-14-0"></span><sup>8.</sup> Our T5 and Switch models were pre-trained with 220 tokens per batch for 550k steps on a revised C4 data set for fair comparisons.
+<span id="page-14-0"></span>8. Our T5 and Switch models were pre-trained with $2^{20}$ tokens per batch for 550k steps on a revised C4 data set for fair comparisons.
 
 | Model        | GLUE        | SQuAD         | SuperGLUE    | Winogrande (XL) |
 |--------------|-------------|---------------|--------------|-----------------|
@@ -273,17 +289,17 @@ Table 5: Fine-tuning results. Fine-tuning results of T5 baselines and Switch mod
 
 Deploying massive neural networks with billions, or trillions, of parameters is inconvenient. To alleviate this, we study distilling [(Hinton et al.,](#page-36-3) [2015)](#page-36-3) large sparse models into small dense models. Future work could additionally study distilling large models into smaller sparse models.
 
-Distillation techniques. In Table [6](#page-16-1) we study a variety of distillation techniques. These techniques are built off of [Sanh et al.](#page-38-8) [(2019)](#page-38-8), who study distillation methods for BERT models. We find that initializing the dense model with the non-expert weights yields a modest improvement. This is possible since all models are FLOP matched, so non-expert layers will have the same dimensions. Since expert layers are usually only added at every or every other FFN layer in a Transformer, this allows for many of the weights to be initialized with trained parameters. Furthermore, we observe a distillation improvement using a mixture of 0.25 for the teacher probabilities and 0.75 for the ground truth label. By combining both techniques we preserve $\approx 30\%$ of the quality gains from the larger sparse models with only $\approx 1/20^{th}$ of the parameters. The quality gain refers to the percent of
+Distillation techniques. In Table [6](#page-16-1) we study a variety of distillation techniques. These techniques are built off of [Sanh et al.](#page-38-8) [(2019)](#page-38-8), who study distillation methods for BERT models. We find that initializing the dense model with the non-expert weights yields a modest improvement. This is possible since all models are FLOP matched, so non-expert layers will have the same dimensions. Since expert layers are usually only added at every or every other FFN layer in a Transformer, this allows for many of the weights to be initialized with trained parameters. Furthermore, we observe a distillation improvement using a mixture of 0.25 for the teacher probabilities and 0.75 for the ground truth label. By combining both techniques we preserve $\approx$ 30% of the quality gains from the larger sparse models with only $\approx$ 1/20th of the parameters. The quality gain refers to the percent of
 
-| Technique                                 | Parameters | Quality (↑)                                  |
-|-------------------------------------------|------------|----------------------------------------------|
-| T5-Base                                   | 223M       | -1.636                                       |
-| Switch-Base                               | 3,800M     | -1.444                                       |
-| Distillation                              | 223M       | <span style="color:blue">(3%)</span> -1.631  |
-| + Init. non-expert weights from teacher   | 223M       | <span style="color:blue">(20%)</span> -1.598 |
-| + 0.75 mix of hard and soft loss          | 223M       | <span style="color:blue">(29%)</span> -1.580 |
-| Initialization Baseline (no distillation) |            |                                              |
-| Init. non-expert weights from teacher     | 223M       | -1.639                                       |
+| Technique                                 | Parameters | Quality (↑)  |
+|-------------------------------------------|------------|--------------|
+| T5-Base                                   | 223M       | -1.636       |
+| Switch-Base                               | 3,800M     | -1.444       |
+| Distillation                              | 223M       | (3%) -1.631  |
+| + Init. non-expert weights from teacher   | 223M       | (20%) -1.598 |
+| + 0.75 mix of hard and soft loss          | 223M       | (29%) -1.580 |
+| Initialization Baseline (no distillation) |            |              |
+| Init. non-expert weights from teacher     | 223M       | -1.639       |
 
 the quality difference between Switch-Base (Teacher) and T5-Base (Student). Therefore, a quality gain of 100% implies the Student equals the performance of the Teacher.
 
@@ -315,7 +331,7 @@ In Figure [7](#page-18-0) we plot the quality improvement in negative log perple
 | Distilled T5-Base | 223M       | 124B  | (30%) 76.6    |
 
 - <span id="page-17-1"></span>Table 8: Distilling a fine-tuned SuperGLUE model. We distill a Switch-Base model finetuned on the SuperGLUE tasks into a T5-Base model. We observe that on smaller data sets our large sparse model can be an effective teacher for distillation. We find that we again achieve 30% of the teacher's performance on a 97% compressed model.
-pre-training both versions for 1M steps, we find that on all 101 languages considered, Switch Transformer increases the final negative log perplexity over the baseline. In Figure [8,](#page-18-1) we present a different view and now histogram the per step speed-up of using Switch Transformer over the mT5-Base.[9](#page-17-2) We find a mean speed-up over mT5-Base of 5x and that 91% of languages achieve at least a 4x speedup. This presents evidence that Switch Transformers are effective multi-task and multi-lingual learners.
+pre-training both versions for 1M steps, we find that on all 101 languages considered, Switch Transformer increases the final negative log perplexity over the baseline. In Figure 8, we present a different view and now histogram the per step speed-up of using Switch Transformer over the mT5-Base.[9](#page-17-2) We find a mean speed-up over mT5-Base of 5x and that 91% of languages achieve at least a 4x speedup. This presents evidence that Switch Transformers are effective multi-task and multi-lingual learners.
 
 ### <span id="page-17-0"></span>5. Designing Models with Data, Model, and Expert-Parallelism
 
@@ -333,18 +349,18 @@ Arbitrarily increasing the number of experts is subject to diminishing returns (
 
 and computation performed and is ultimately limited by the memory per accelerator. Once it exceeds the size of the accelerator's memory, single program multiple data (SPMD) modelparallelism can be employed. This section studies the trade-offs of combining data, model, and expert-parallelism.
 
-Reviewing the Feed-Forward Network (FFN) Layer. We use the FFN layer as an example of how data, model and expert-parallelism works in Mesh TensorFlow [(Shazeer](#page-38-3) [et al.,](#page-38-3) [2018)](#page-38-3) and review it briefly here. We assume B tokens in the batch, each of dimension $d_{model}$. Both the input (x) and output (y) of the FFN are of size [$B, d_{model}$] and the intermediate (h) is of size [$B, d_{ff}$] where $d_{ff}$ is typically several times larger than $d_{model}$. In the FFN, the intermediate is $h = xW_{in}$ and then the output of the layer is $y = ReLU(h)W_{out}$. Thus $W_{in}$ and $W_{out}$ are applied independently to each token and have sizes [$d_{model}, d_{ff}$] and [$d_{ff}, d_{model}$].
+Reviewing the Feed-Forward Network (FFN) Layer. We use the FFN layer as an example of how data, model and expert-parallelism works in Mesh TensorFlow [(Shazeer](#page-38-3) [et al.,](#page-38-3) [2018)](#page-38-3) and review it briefly here. We assume $B$ tokens in the batch, each of dimension $d_{model}$. Both the input $(x)$ and output $(y)$ of the FFN are of size $[B, d_{model}]$ and the intermediate $(h)$ is of size $[B, d_{ff}]$ where $d_{ff}$ is typically several times larger than $d_{model}$. In the FFN, the intermediate is $h = xW_{in}$ and then the output of the layer is $y = ReLU(h)W_{out}$. Thus $W_{in}$ and $W_{out}$ are applied independently to each token and have sizes $[d_{model}, d_{ff}] $and $[d_{ff}, d_{model}]$.
 
-We describe two aspects of partitioning: how the weights and batches of data divide over cores, depicted in Figure [9.](#page-20-1) We denote all cores available as N which Mesh Tensorflow may then remap into a logical multidimensional mesh of processors. Here we create a two-dimensional logical mesh, with one dimension representing the number of ways for data-parallel sharding ($n$) and the other, the model-parallel sharding ($m$). The total cores must equal the ways to shard across both data and model-parallelism, e.g. $N = n \times m$. To shard the layer across cores, the tensors containing that batch of B tokens are sharded across $n$ data-parallel cores, so each core contains $B/n$ tokens. Tensors and variables with $d_{ff}$ are then sharded across $m$ model-parallel cores. For the variants with experts-layers, we consider E experts, each of which can process up to C tokens.
+We describe two aspects of partitioning: how the weights and batches of data divide over cores, depicted in Figure [9.](#page-20-1) We denote all cores available as $N$ which Mesh Tensorflow may then remap into a logical multidimensional mesh of processors. Here we create a two-dimensional logical mesh, with one dimension representing the number of ways for data-parallel sharding $(n)$ and the other, the model-parallel sharding $(m)$. The total cores must equal the ways to shard across both data and model-parallelism, e.g. $N = n \times m$. To shard the layer across cores, the tensors containing that batch of $B$ tokens are sharded across $n$ data-parallel cores, so each core contains $B/n$ tokens. Tensors and variables with $d_{ff}$ are then sharded across $m$ model-parallel cores. For the variants with experts-layers, we consider $E$ experts, each of which can process up to $C$ tokens.
 
 | Term | Description                                     |
 |------|-------------------------------------------------|
-| $B$  | Number of tokens in the batch.                  |
-| $N$  | Number of total cores.                          |
-| $n$  | Number of ways for data-parallelism sharding.   |
-| $m$  | Number of ways for model-parallelism sharding.  |
-| $E$  | Number of experts in Switch layers.             |
-| $C$  | Expert capacity, the batch size of each expert. |
+| B    | Number of tokens in the batch.                  |
+| N    | Number of total cores.                          |
+| n    | Number of ways for data-parallelism sharding.   |
+| m    | Number of ways for model-parallelism sharding.  |
+| E    | Number of experts in Switch layers.             |
+| C    | Expert capacity, the batch size of each expert. |
 
 #### <span id="page-19-0"></span>5.1 Data Parallelism
 
@@ -352,7 +368,7 @@ When training data parallel models, which is the standard for distributed traini
 
 #### <span id="page-19-1"></span>5.2 Model Parallelism
 
-We now consider a scenario where all cores are allocated exclusively to the model-parallel dimension and so $n = 1, m = N$. Now all cores must keep the full $B$ tokens and each core will contain a unique slice of the weights. For each forward and backward pass, a communication cost is now incurred. Each core sends a tensor of [$B, d_{model}$] to compute the second matrix multiplication $ReLU(h)W_{out}$ because the $d_{ff}$ dimension is partitioned and must be summed over. As a general rule, whenever a dimension that is partitioned across cores must be summed, then an all-reduce operation is added for both the forward and backward pass. This contrasts with pure data parallelism where an all-reduce only occurs at the end of the entire forward and backward pass.
+We now consider a scenario where all cores are allocated exclusively to the model-parallel dimension and so $n = 1, m = N$. Now all cores must keep the full $B$ tokens and each core will contain a unique slice of the weights. For each forward and backward pass, a communication cost is now incurred. Each core sends a tensor of $[B, d_{model}]$ to compute the second matrix multiplication $ReLU(h)W_{out}$ because the $d_{ff}$ dimension is partitioned and must be summed over. As a general rule, whenever a dimension that is partitioned across cores must be summed, then an all-reduce operation is added for both the forward and backward pass. This contrasts with pure data parallelism where an all-reduce only occurs at the end of the entire forward and backward pass.
 
 ![](_page_20_Figure_1.jpeg)
 
@@ -369,41 +385,42 @@ It is common to mix both model and data parallelism for large scale models, whic
 
 #### <span id="page-21-0"></span>5.4 Expert and Data Parallelism
 
-Next we describe the partitioning strategy for expert and data parallelism. Switch Transformers will allocate all of their cores to the data partitioning dimension n, which will also correspond to the number of experts in the model. For each token per core a router locally computes assignments to the experts. The output is a binary matrix of size $[n, B/n, E$, C] which is partitioned across the first dimension and determines expert assignment. This binary matrix is then used to do a gather via matrix multiplication with the input tensor of $[n, B/n, d_{model}]$.
+Next we describe the partitioning strategy for expert and data parallelism. Switch Transformers will allocate all of their cores to the data partitioning dimension $n$, which will also correspond to the number of experts in the model. For each token per core a router locally computes assignments to the experts. The output is a binary matrix of size $[n, B/n, E$, $C]$ which is partitioned across the first dimension and determines expert assignment. This binary matrix is then used to do a gather via matrix multiplication with the input tensor of $[n, B/n, d_{model}]$.
 
 $$\text{einsum}([n, B/n, d_{model}], [n, B/n, E, C], \text{dimension} = [B/n]) \tag{7}$$
 
-resulting in the final tensor of shape $[n, E, C, d_{model}]$, which is sharded across the first dimension. Because each core has its own expert, we do an all-to-all communication of size $[E, C, d_{model}]$ to now shard the $E$ dimension instead of the $n$-dimension. There are additional communication costs of bfloat16 tensors of size $E \times C \times d_{model}$ in the forward pass to analogusly receive the tokens from each expert located on different cores. See Appendix [F](#page-32-0) for a detailed analysis of the expert partitioning code.
+resulting in the final tensor of shape $[n, E, C, d_{model}]$, which is sharded across the first dimension. Because each core has its own expert, we do an all-to-all communication of size $[E, C, d_{model}]$ to now shard the $E$ dimension instead of the $n$-dimension. There are additional communication costs of bfloat16 tensors of size $E \times C \times d_{model}$ in the forward pass to analogously receive the tokens from each expert located on different cores. See Appendix [F](#page-32-0) for a detailed analysis of the expert partitioning code.
 
 #### <span id="page-21-1"></span>5.5 Expert, Model and Data Parallelism
 
-In the design of our best model, we seek to balance the FLOPS per token and the parameter count. When we scale the number of experts, we increase the number of parameters, but do not change the FLOPs per token. In order to increase FLOPs, we must also increase the $d_{ff}$ dimension (which also increases parameters, but at a slower rate). This presents a trade-off: as we increase $d_{ff}$ we will run out of memory per core, which then necessitates increasing $m$. But since we have a fixed number of cores $N$, and $N = n \times m$, we must decrease $n$, which forces use of a smaller batch-size (in order to hold tokens per core constant).
+In the design of our best model, we seek to balance the FLOPS per token and the parameter count. When we scale the number of experts, we increase the number of parameters, but do not change the FLOPS per token. In order to increase FLOPS, we must also increase the $d_{ff}$ dimension (which also increases parameters, but at a slower rate). This presents a trade-off: as we increase $d_{ff}$ we will run out of memory per core, which then necessitates increasing $m$. But since we have a fixed number of cores $N$, and $N = n \times m$, we must decrease $n$, which forces use of a smaller batch-size (in order to hold tokens per core constant).
 
 When combining both model and expert-parallelism, we will have all-to-all communication costs from routing the tokens to the correct experts along with the internal all-reduce communications from the model parallelism. Balancing the FLOPS, communication costs and memory per core becomes quite complex when combining all three methods where the best mapping is empirically determined. See our further analysis in section [5.6](#page-21-2) for how the number of experts effects the downstream performance as well.
 
 #### <span id="page-21-2"></span>5.6 Towards Trillion Parameter Models
 
-Combining expert, model and data parallelism, we design two large Switch Transformer models, one with 395 billion and 1.6 trillion parameters, respectively. We study how these models perform on both up-stream pre-training as language models and their downstream fine-tuning performance. The parameters, FLOPs per sequence and hyper-parameters of the two different models are listed below in Table [9.](#page-22-0) Standard hyper-parameters of the Transformer, including $d_{model}$, $d_{ff}$, $d_{kv}$, number of heads and number of layers are described, as well as a less common feature, $FFN_{GEGLU}$, which refers to a variation of the FFN layer where the expansion matrix is substituted with two sets of weights which are non-linearly combined [(Shazeer,](#page-38-9) [2020)](#page-38-9).
+Combining expert, model and data parallelism, we design two large Switch Transformer models, one with 395 billion and 1.6 trillion parameters, respectively. We study how these models perform on both up-stream pre-training as language models and their downstream fine-tuning performance. The parameters, FLOPs per sequence and hyper-parameters of the two different models are listed below in Table 9. Standard hyper-parameters of the Transformer, including $d_{model}$, $d_{ff}$, $d_{kv}$, number of heads and number of layers are described, as well as a less common feature, $FFN_{GEGLU}$, which refers to a variation of the FFN layer where the expansion matrix is substituted with two sets of weights which are non-linearly combined (Shazeer, 2020).
 
 The Switch-C model is designed using only expert-parallelism, and no model-parallelism, as described earlier in Section [5.4.](#page-21-0) As a result, the hyper-parameters controlling the width,
 
-| Model        | Parameters   | FLOPs/seq   | $d_{model}$ | FFN $_{GEGLU}$       | $d_{ff}$              | $d_{kv}$ | Num. Heads |
-|--------------|--------------|-------------|-------------|----------------------|-----------------------|----------|------------|
-| T5-Base      | 0.2B         | 124B        | 768         | ✓                    | 2048                  | 64       | 12         |
-| T5-Large     | 0.7B         | 425B        | 1024        | ✓                    | 2816                  | 64       | 16         |
-| T5-XXL       | 11B          | 6.3T        | 4096        | ✓                    | 10240                 | 64       | 64         |
-| Switch-Base  | 7B           | 124B        | 768         | ✓                    | 2048                  | 64       | 12         |
-| Switch-Large | 26B          | 425B        | 1024        | ✓                    | 2816                  | 64       | 16         |
-| Switch-XXL   | 395B         | 6.3T        | 4096        | ✓                    | 10240                 | 64       | 64         |
-| Switch-C     | 1571B        | 890B        | 2080        | ✓                    | 6144                  | 64       | 32         |
-| Model        | Expert Freq. | Num. Layers | Num Experts | Neg. Log Perp. @250k | Neg. Log Perp. @ 500k |          |            |
-| T5-Base      | –            | 12          | –           | -1.599               | -1.556                |          |            |
-| T5-Large     | –            | 24          | –           | -1.402               | -1.350                |          |            |
-| T5-XXL       | –            | 24          | –           | -1.147               | -1.095                |          |            |
-| Switch-Base  | 1/2          | 12          | 128         | -1.370               | -1.306                |          |            |
-| Switch-Large | 1/2          | 24          | 128         | -1.248               | -1.177                |          |            |
-| Switch-XXL   | 1/2          | 24          | 64          | -1.086               | -1.008                |          |            |
-| Switch-C     | 1            | 15          | 2048        | -1.096               | -1.043                |          |            |
+| Model        | Parameters   | FLOPs/seq   | dmodel      | F F NGEGLU           | df f                  | dkv | Num. Heads |
+|--------------|--------------|-------------|-------------|----------------------|-----------------------|-----|------------|
+| T5-Base      | 0.2B         | 124B        | 768         | X                    | 2048                  | 64  | 12         |
+| T5-Large     | 0.7B         | 425B        | 1024        | X                    | 2816                  | 64  | 16         |
+| T5-XXL       | 11B          | 6.3T        | 4096        | X                    | 10240                 | 64  | 64         |
+| Switch-Base  | 7B           | 124B        | 768         | X                    | 2048                  | 64  | 12         |
+| Switch-Large | 26B          | 425B        | 1024        | X                    | 2816                  | 64  | 16         |
+| Switch-XXL   | 395B         | 6.3T        | 4096        | X                    | 10240                 | 64  | 64         |
+| Switch-C     | 1571B        | 890B        | 2080        |                      | 6144                  | 64  | 32         |
+|              |              |             |             |                      |                       |     |            |
+| Model        | Expert Freq. | Num. Layers | Num Experts | Neg. Log Perp. @250k | Neg. Log Perp. @ 500k |     |            |
+| T5-Base      | –            | 12          | –           | -1.599               | -1.556                |     |            |
+| T5-Large     | –            | 24          | –           | -1.402               | -1.350                |     |            |
+| T5-XXL       | –            | 24          | –           | -1.147               | -1.095                |     |            |
+| Switch-Base  | 1/2          | 12          | 128         | -1.370               | -1.306                |     |            |
+| Switch-Large | 1/2          | 24          | 128         | -1.248               | -1.177                |     |            |
+| Switch-XXL   | 1/2          | 24          | 64          | -1.086               | -1.008                |     |            |
+| Switch-C     | 1            | 15          | 2048        | -1.096               | -1.043                |     |            |
 
 - <span id="page-22-0"></span>Table 9: Switch model design and pre-training performance. We compare the hyperparameters and pre-training performance of the T5 models to our Switch Transformer variants. The last two columns record the pre-training model quality on the C4 data set after 250k and 500k steps, respectively. We observe that the Switch-C Transformer variant is 4x faster to a fixed perplexity (with the same compute budget) than the T5-XXL model, with the gap increasing as training progresses.
 depth, number of heads, and so on, are all much smaller than the T5-XXL model. In contrast, the Switch-XXL is FLOP-matched to the T5-XXL model, which allows for larger dimensions of the hyper-parameters, but at the expense of additional communication costs induced by model-parallelism (see Section [5.5](#page-21-1) for more details).
@@ -428,7 +445,7 @@ Our work studies a specific model in a class of methods that do conditional comp
 
 Mixture of Experts (MoE), in the context of modern deep learning architectures, was proven effective in [Shazeer et al.](#page-38-2) [(2017)](#page-38-2). That work added an MoE layer which was stacked between LSTM [(Hochreiter and Schmidhuber,](#page-36-10) [1997)](#page-36-10) layers, and tokens were separately routed to combinations of experts. This resulted in state-of-the-art results in language modeling and machine translation benchmarks. The MoE layer was reintroduced into the Transformer architecture by the Mesh Tensorflow library [(Shazeer et al.,](#page-38-3) [2018)](#page-38-3) where MoE layers were introduced as a substitute of the FFN layers, however, there were no accompanying NLP results. More recently, through advances in machine learning infrastructure, GShard [(Lepikhin et al.,](#page-37-2) [2020)](#page-37-2), which extended the XLA compiler, used the MoE Transformer to dramatically improve machine translation across 100 languages. Finally [Fan et al.](#page-35-10) [(2021)](#page-35-10) chooses a different deterministic MoE strategy to split the model parameters into non-overlapping groups of languages.
 
-Sparsity along the sequence length dimension (L) in the Transformer attention patterns has been a successful technique to reduce the attention complexity from $O(L^2)$ (Child et al., [2019;](#page-35-11) [Correia et al.,](#page-35-12) [2019;](#page-35-12) [Sukhbaatar et al.,](#page-38-11) [2019;](#page-38-11) [Kitaev et al.,](#page-36-11) [2020;](#page-36-11) [Zaheer et al.,](#page-39-7) [2020;](#page-39-7) [Beltagy et al.,](#page-35-13) [2020)](#page-35-13). This has enabled learning longer sequences than previously possible. This version of the Switch Transformer does not employ attention sparsity, but these techniques are complimentary, and, as future work, these could be combined to potentially improve learning on tasks requiring long contexts.
+Sparsity along the sequence length dimension $(L)$ in the Transformer attention patterns has been a successful technique to reduce the attention complexity from $O(L^2)$ (Child et al., 2019; Correia et al., 2019; Sukhbaatar et al., 2019; Kitaev et al., 2020; Zaheer et al., 2020; Beltagy et al., 2020). This has enabled learning longer sequences than previously possible. This version of the Switch Transformer does not employ attention sparsity, but these techniques are complimentary, and, as future work, these could be combined to potentially improve learning on tasks requiring long contexts.
 
 ### <span id="page-24-0"></span>7. Discussion
 
@@ -440,9 +457,9 @@ I don't have access to a supercomputer—is this still useful for me? Though thi
 
 Do sparse models outperform dense models on the speed-accuracy Pareto curve? Yes. Across a wide variety of different models sizes, sparse models outperform dense models per step and on wall clock time. Our controlled experiments show for a fixed amount of computation and time, sparse models outperform dense models.
 
-I can't deploy a trillion parameter model—can we shrink these models? We cannot fully preserve the model quality, but compression rates of 10 to 100x are achievable by distilling our sparse models into dense models while achieving ≈30% of the quality gain of the expert model.
+I can't deploy a trillion parameter model—can we shrink these models? We cannot fully preserve the model quality, but compression rates of 10 to 100x are achievable by distilling our sparse models into dense models while achieving $\approx$30% of the quality gain of the expert model.
 
-Why use Switch Transformer instead of a model-parallel dense model? On a time basis, Switch Transformers can be far more efficient than dense-models with sharded parameters (Figure [6)](#page-13-2). Also, we point out that this decision is not mutually exclusive—we can, and do, use model-parallelism in Switch Transformers, increasing the FLOPs per token, but incurring the slowdown of conventional model-parallelism.
+Why use Switch Transformer instead of a model-parallel dense model? On a time basis, Switch Transformers can be far more efficient than dense-models with sharded parameters (Figure [6)](#page-13-2). Also, we point out that this decision is *not* mutually exclusive—we can, and do, use model-parallelism in Switch Transformers, increasing the FLOPs per token, but incurring the slowdown of conventional model-parallelism.
 
 Why aren't sparse models widely used already? The motivation to try sparse models has been stymied by the massive success of scaling dense models (the success of which is partially driven by co-adaptation with deep learning hardware as argued in [Hooker](#page-36-12) [(2020)](#page-36-12)). Further, sparse models have been subject to multiple issues including (1) model complexity, (2) training difficulties, and (3) communication costs. Switch Transformer makes strides to alleviate these issues.
 
@@ -496,7 +513,7 @@ Table [10](#page-27-1) records the quality after a fixed number of steps as well
 
 Due to software constraints on TPU accelerators, the shapes of our Tensors must be statically sized. As a result, each expert has a finite and fixed capacity to process token representations. This, however, presents an issue for our model which dynamically routes tokens at run-time that may result in an uneven distribution over experts. If the number of tokens sent to an expert is less than the expert capacity, then the computation may simply be padded – an inefficient use of the hardware, but mathematically correct. However, when the number of tokens sent to an expert is larger than its capacity (expert overflow), a protocol is needed to handle this. [Lepikhin et al.](#page-37-2) [(2020)](#page-37-2) adapts a Mixture-of-Expert model and addresses expert overflow by passing its representation to the next layer without processing through a residual connection which we also follow.
 
-We suspected that having no computation applied to tokens could be very wasteful, especially since if there is overflow on one expert, that means another expert will have extra capacity. With this intuition we create No-Token-Left-Behind, which iteratively reroutes any tokens that are at first routed to an expert that is overflowing. Figure [11](#page-29-0) shows a graphical description of this method, which will allow us to guarantee almost no tokens will be dropped during training and inference. We hypothesised that this could improve performance and further stabilize training, but we found no empirical benefits. We suspect that once the network learns associations between different tokens and experts, if this association is changed (e.g. sending a token to its second highest expert) then performance could be degraded.
+We suspected that having no computation applied to tokens could be very wasteful, especially since if there is overflow on one expert, that means another expert will have extra capacity. With this intuition we create *No-Token-Left-Behind*, which iteratively reroutes any tokens that are at first routed to an expert that is overflowing. Figure [11](#page-29-0) shows a graphical description of this method, which will allow us to guarantee almost no tokens will be dropped during training and inference. We hypothesised that this could improve performance and further stabilize training, but we found no empirical benefits. We suspect that once the network learns associations between different tokens and experts, if this association is changed (e.g. sending a token to its second highest expert) then performance could be degraded.
 
 ### <span id="page-28-1"></span>C. Encouraging Exploration Across Experts
 
@@ -517,7 +534,7 @@ Switch Transformer is also an effective architecture at small scales as well as
 | Argmax         | -1.471                       |
 | Sample softmax | -1.570                       |
 | Input dropout  | -1.480                       |
-| Input jitter   | -1.468                       |
+| Input jitter   | <b>-1.468</b>                |
 
 - <span id="page-29-1"></span>Table 11: Router Exploration Strategies. Quality of the Switch Transformer, measured by the negative log perplexity, under different randomness-strategies for selecting the expert (lower is better). There is no material speed performance difference between the variants.
 at the scale of 10B+ parameter models, but we show in Figure [12](#page-30-0) as few as 2 experts produce compelling gains over a FLOP-matched counterpart. Even if a super computer is not readily available, training Switch Transformers with 2, 4, or 8 experts (as we typically recommend one expert per core) results in solid improvements over T5 dense baselines.
@@ -585,6 +602,11 @@ router logits = mtf.einsum([inputs, router weights], reduced dim=d model)
 
 # Add noise for exploration across experts. router logits += mtf.random uniform(shape=router logits.shape, minval=1−eps, maxval=1+eps)
 
+```
+# Convert input to softmax operation from bfloat16 to float32 for stability.
+router logits = mtf.to float32(router logits)
+```
+
 ```
 # Probabilities for each token of what expert it should be sent to.
 router probs = mtf.softmax(router logits, axis=−1)
diff --git a/marker/processors/debug.py b/marker/processors/debug.py
index 90dea5df..cf9f0da6 100644
--- a/marker/processors/debug.py
+++ b/marker/processors/debug.py
@@ -69,6 +69,7 @@ def draw_pdf_debug_images(self, document: Document):
                 # Skip any blocks that have been removed
                 if child.removed:
                     continue
+
                 if child.block_type == BlockTypes.Line:
                     bbox = child.polygon.rescale(page.polygon.size, png_image.size).bbox
                     line_bboxes.append(bbox)
@@ -78,7 +79,6 @@ def draw_pdf_debug_images(self, document: Document):
                     span_bboxes.append(bbox)
 
             self.render_on_image(line_bboxes, png_image, color="blue", draw_bbox=True, label_font_size=24, labels=[str(i) for i in line_ids])
-            #self.render_on_image(span_bboxes, png_image, color="green", draw_bbox=True, label_font_size=24)
 
             png_image = self.render_layout_boxes(page, png_image)
 
diff --git a/marker/processors/line_merge.py b/marker/processors/line_merge.py
index 2ba2b42b..9c45bde9 100644
--- a/marker/processors/line_merge.py
+++ b/marker/processors/line_merge.py
@@ -1,7 +1,8 @@
-from typing import Annotated
+from typing import Annotated, List
 
 from marker.processors import BaseProcessor
 from marker.schema import BlockTypes
+from marker.schema.blocks import Block
 from marker.schema.document import Document
 from marker.schema.text import Line
 from marker.util import matrix_intersection_area
@@ -11,11 +12,15 @@ class LineMergeProcessor(BaseProcessor):
     """
     A processor for merging inline math lines.
     """
-    block_types = (BlockTypes.Text, BlockTypes.TextInlineMath)
+    block_types = (BlockTypes.Text, BlockTypes.TextInlineMath, BlockTypes.Caption, BlockTypes.Footnote, BlockTypes.SectionHeader)
     min_merge_pct: Annotated[
         float,
         "The minimum percentage of intersection area to consider merging."
-    ] = .02
+    ] = .015
+    block_expand_threshold: Annotated[
+        float,
+        "The percentage of the block width to expand the bounding box."
+    ] = .05
     min_merge_ydist: Annotated[
         float,
         "The minimum y distance between lines to consider merging."
@@ -23,12 +28,96 @@ class LineMergeProcessor(BaseProcessor):
     intersection_pct_threshold: Annotated[
         float,
         "The total amount of intersection area concentrated in the max intersection block."
-    ] = .7
+    ] = .6
+    vertical_overlap_pct_threshold: Annotated[
+        float,
+        "The minimum percentage of vertical overlap to consider merging."
+    ] = .8
+    use_llm: Annotated[
+        bool,
+        "Whether to use LLMs to improve accuracy."
+    ] = False
 
     def __init__(self, config):
         super().__init__(config)
 
+    def merge_lines(self, lines: List[Line], block: Block):
+        lines = [l for l in lines if l.polygon.width * 5 > l.polygon.height]  # Skip vertical lines
+        line_bboxes = [l.polygon.expand(self.block_expand_threshold, 0).bbox for l in lines]  # Expand horizontally
+        intersections = matrix_intersection_area(line_bboxes, line_bboxes)
+
+        merges = []
+        merge = []
+        for i in range(len(line_bboxes)):
+            intersection_row = intersections[i]
+            intersection_row[i] = 0  # Zero out the current idx
+
+            if i < len(line_bboxes) - 1:
+                intersection_row[i+1] = 0 # Zero out the next idx, so we only evaluate merge from the left
+
+            if len(merge) == 0:
+                merge.append(i)
+                continue
+
+            # Zero out previous merge segments
+            merge_intersection = sum([intersection_row[m] for m in merge])
+            line_area = lines[i].polygon.area
+            intersection_pct = merge_intersection / max(1, line_area)
+
+            total_intersection = max(1, sum(intersection_row))
+
+            line_start = lines[merge[0]].polygon.y_start
+            line_end = lines[merge[0]].polygon.y_end
+
+            vertical_overlap_start = max(line_start, lines[i].polygon.y_start)
+            vertical_overlap_end = min(line_end, lines[i].polygon.y_end)
+            vertical_overlap = max(0, vertical_overlap_end - vertical_overlap_start)
+            vertical_overlap_pct = vertical_overlap / max(1, lines[i].polygon.height)
+
+            if all([
+                # Overlaps enough
+                intersection_pct >= self.min_merge_pct,
+                # Within same line
+                vertical_overlap_pct > self.vertical_overlap_pct_threshold,
+                # doesn't overlap with anything else
+                merge_intersection / total_intersection >= self.intersection_pct_threshold
+            ]):
+                merge.append(i)
+            else:
+                merges.append(merge)
+                merge = []
+
+        if merge:
+            merges.append(merge)
+
+        merges = [m for m in merges if len(m) > 1]
+        merged = set()
+        for merge in merges:
+            merge = [m for m in merge if m not in merged]
+            if len(merge) < 2:
+                continue
+
+            line: Line = lines[merge[0]]
+            merged.add(merge[0])
+            for idx in merge[1:]:
+                other_line: Line = lines[idx]
+                line.merge(other_line)
+                block.structure.remove(other_line.id)
+                other_line.removed = True  # Mark line as removed
+                merged.add(idx)
+
+            # It is probably math if we are merging provider lines like this
+            if not line.formats:
+                line.formats = ["math"]
+            elif "math" not in line.formats:
+                line.formats.append("math")
+
+
     def __call__(self, document: Document):
+        # Merging lines only needed for inline math
+        if not self.use_llm:
+            return
+
         for page in document.pages:
             for block in page.contained_blocks(document, self.block_types):
                 if block.structure is None:
@@ -38,62 +127,4 @@ def __call__(self, document: Document):
                     continue
 
                 lines = block.contained_blocks(document, (BlockTypes.Line,))
-                lines = [l for l in lines if l.polygon.width * 5 > l.polygon.height]  # Skip vertical lines
-                line_bboxes = [l.polygon.expand(self.min_merge_pct, 0).bbox for l in lines] # Expand horizontally
-                intersections = matrix_intersection_area(line_bboxes, line_bboxes)
-
-                merges = []
-                merge = []
-                for i in range(len(line_bboxes) - 1):
-                    next_idx = i + 1
-                    intersection_val = intersections[i, next_idx]
-                    intersection_pct = intersection_val / max(1, lines[i].polygon.area)
-                    intersection_row = intersections[i]
-                    intersection_row[i] = 0 # Zero out the current idx
-
-                    # Zero out previous merge segments
-                    for m in merge:
-                        intersection_row[m] = 0
-                    max_intersection_idx = intersection_row.argmax()
-                    total_intersection = max(1, sum(intersection_row))
-                    max_intersection = intersection_row[max_intersection_idx]
-
-
-                    if all([
-                        max_intersection_idx == next_idx, # The next line is the max intersection line
-                        intersection_pct >= self.min_merge_pct,
-                        abs(lines[i].polygon.y_start - lines[next_idx].polygon.y_start) <= self.min_merge_ydist,
-                        abs(lines[i].polygon.y_end - lines[next_idx].polygon.y_end) <= self.min_merge_ydist,
-                        max_intersection / total_intersection >= self.intersection_pct_threshold
-                    ]):
-                        if not merge:
-                            merge.append(i)
-                        merge.append(next_idx)
-                    else:
-                        merges.append(merge)
-                        merge = []
-
-                if merge:
-                    merges.append(merge)
-
-                merges = [m for m in merges if len(m) > 1]
-                merged = set()
-                for merge in merges:
-                    merge = [m for m in merge if m not in merged]
-                    if len(merge) < 2:
-                        continue
-
-                    line: Line = lines[merge[0]]
-                    merged.add(merge[0])
-                    for idx in merge[1:]:
-                        other_line: Line = lines[idx]
-                        line.merge(other_line)
-                        block.structure.remove(other_line.id)
-                        other_line.removed = True # Mark line as removed
-                        merged.add(idx)
-
-                    # It is probably math if we are merging provider lines like this
-                    if not line.formats:
-                        line.formats = ["math"]
-                    elif "math" not in line.formats:
-                        line.formats.append("math")
+                self.merge_lines(lines, block)
diff --git a/marker/renderers/markdown.py b/marker/renderers/markdown.py
index 3d10af9f..debac078 100644
--- a/marker/renderers/markdown.py
+++ b/marker/renderers/markdown.py
@@ -172,7 +172,8 @@ def convert_table(self, el, text, convert_as_inline):
 
     def convert_a(self, el, text, convert_as_inline):
         text = self.escape(text)
-        text = re.sub(r"([\[\]])", r"\\\1", text)
+        # Escape brackets and parentheses in text
+        text = re.sub(r"([\[\]()])", r"\\\1", text)
         return super().convert_a(el, text, convert_as_inline)
 
     def convert_span(self, el, text, convert_as_inline):
diff --git a/poetry.lock b/poetry.lock
index 4365d517..f1e90012 100644
--- a/poetry.lock
+++ b/poetry.lock
@@ -13,92 +13,92 @@ files = [
 
 [[package]]
 name = "aiohttp"
-version = "3.11.12"
+version = "3.11.13"
 description = "Async http client/server framework (asyncio)"
 optional = false
 python-versions = ">=3.9"
 files = [
-    {file = "aiohttp-3.11.12-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:aa8a8caca81c0a3e765f19c6953416c58e2f4cc1b84829af01dd1c771bb2f91f"},
-    {file = "aiohttp-3.11.12-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:84ede78acde96ca57f6cf8ccb8a13fbaf569f6011b9a52f870c662d4dc8cd854"},
-    {file = "aiohttp-3.11.12-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:584096938a001378484aa4ee54e05dc79c7b9dd933e271c744a97b3b6f644957"},
-    {file = "aiohttp-3.11.12-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:392432a2dde22b86f70dd4a0e9671a349446c93965f261dbaecfaf28813e5c42"},
-    {file = "aiohttp-3.11.12-cp310-cp310-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:88d385b8e7f3a870146bf5ea31786ef7463e99eb59e31db56e2315535d811f55"},
-    {file = "aiohttp-3.11.12-cp310-cp310-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:b10a47e5390c4b30a0d58ee12581003be52eedd506862ab7f97da7a66805befb"},
-    {file = "aiohttp-3.11.12-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:0b5263dcede17b6b0c41ef0c3ccce847d82a7da98709e75cf7efde3e9e3b5cae"},
-    {file = "aiohttp-3.11.12-cp310-cp310-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:50c5c7b8aa5443304c55c262c5693b108c35a3b61ef961f1e782dd52a2f559c7"},
-    {file = "aiohttp-3.11.12-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:d1c031a7572f62f66f1257db37ddab4cb98bfaf9b9434a3b4840bf3560f5e788"},
-    {file = "aiohttp-3.11.12-cp310-cp310-musllinux_1_2_armv7l.whl", hash = "sha256:7e44eba534381dd2687be50cbd5f2daded21575242ecfdaf86bbeecbc38dae8e"},
-    {file = "aiohttp-3.11.12-cp310-cp310-musllinux_1_2_i686.whl", hash = "sha256:145a73850926018ec1681e734cedcf2716d6a8697d90da11284043b745c286d5"},
-    {file = "aiohttp-3.11.12-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:2c311e2f63e42c1bf86361d11e2c4a59f25d9e7aabdbdf53dc38b885c5435cdb"},
-    {file = "aiohttp-3.11.12-cp310-cp310-musllinux_1_2_s390x.whl", hash = "sha256:ea756b5a7bac046d202a9a3889b9a92219f885481d78cd318db85b15cc0b7bcf"},
-    {file = "aiohttp-3.11.12-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:526c900397f3bbc2db9cb360ce9c35134c908961cdd0ac25b1ae6ffcaa2507ff"},
-    {file = "aiohttp-3.11.12-cp310-cp310-win32.whl", hash = "sha256:b8d3bb96c147b39c02d3db086899679f31958c5d81c494ef0fc9ef5bb1359b3d"},
-    {file = "aiohttp-3.11.12-cp310-cp310-win_amd64.whl", hash = "sha256:7fe3d65279bfbee8de0fb4f8c17fc4e893eed2dba21b2f680e930cc2b09075c5"},
-    {file = "aiohttp-3.11.12-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:87a2e00bf17da098d90d4145375f1d985a81605267e7f9377ff94e55c5d769eb"},
-    {file = "aiohttp-3.11.12-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:b34508f1cd928ce915ed09682d11307ba4b37d0708d1f28e5774c07a7674cac9"},
-    {file = "aiohttp-3.11.12-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:936d8a4f0f7081327014742cd51d320296b56aa6d324461a13724ab05f4b2933"},
-    {file = "aiohttp-3.11.12-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:2de1378f72def7dfb5dbd73d86c19eda0ea7b0a6873910cc37d57e80f10d64e1"},
-    {file = "aiohttp-3.11.12-cp311-cp311-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:b9d45dbb3aaec05cf01525ee1a7ac72de46a8c425cb75c003acd29f76b1ffe94"},
-    {file = "aiohttp-3.11.12-cp311-cp311-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:930ffa1925393381e1e0a9b82137fa7b34c92a019b521cf9f41263976666a0d6"},
-    {file = "aiohttp-3.11.12-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:8340def6737118f5429a5df4e88f440746b791f8f1c4ce4ad8a595f42c980bd5"},
-    {file = "aiohttp-3.11.12-cp311-cp311-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:4016e383f91f2814e48ed61e6bda7d24c4d7f2402c75dd28f7e1027ae44ea204"},
-    {file = "aiohttp-3.11.12-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:3c0600bcc1adfaaac321422d615939ef300df81e165f6522ad096b73439c0f58"},
-    {file = "aiohttp-3.11.12-cp311-cp311-musllinux_1_2_armv7l.whl", hash = "sha256:0450ada317a65383b7cce9576096150fdb97396dcfe559109b403c7242faffef"},
-    {file = "aiohttp-3.11.12-cp311-cp311-musllinux_1_2_i686.whl", hash = "sha256:850ff6155371fd802a280f8d369d4e15d69434651b844bde566ce97ee2277420"},
-    {file = "aiohttp-3.11.12-cp311-cp311-musllinux_1_2_ppc64le.whl", hash = "sha256:8fd12d0f989c6099e7b0f30dc6e0d1e05499f3337461f0b2b0dadea6c64b89df"},
-    {file = "aiohttp-3.11.12-cp311-cp311-musllinux_1_2_s390x.whl", hash = "sha256:76719dd521c20a58a6c256d058547b3a9595d1d885b830013366e27011ffe804"},
-    {file = "aiohttp-3.11.12-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:97fe431f2ed646a3b56142fc81d238abcbaff08548d6912acb0b19a0cadc146b"},
-    {file = "aiohttp-3.11.12-cp311-cp311-win32.whl", hash = "sha256:e10c440d142fa8b32cfdb194caf60ceeceb3e49807072e0dc3a8887ea80e8c16"},
-    {file = "aiohttp-3.11.12-cp311-cp311-win_amd64.whl", hash = "sha256:246067ba0cf5560cf42e775069c5d80a8989d14a7ded21af529a4e10e3e0f0e6"},
-    {file = "aiohttp-3.11.12-cp312-cp312-macosx_10_13_universal2.whl", hash = "sha256:e392804a38353900c3fd8b7cacbea5132888f7129f8e241915e90b85f00e3250"},
-    {file = "aiohttp-3.11.12-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:8fa1510b96c08aaad49303ab11f8803787c99222288f310a62f493faf883ede1"},
-    {file = "aiohttp-3.11.12-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:dc065a4285307607df3f3686363e7f8bdd0d8ab35f12226362a847731516e42c"},
-    {file = "aiohttp-3.11.12-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:cddb31f8474695cd61fc9455c644fc1606c164b93bff2490390d90464b4655df"},
-    {file = "aiohttp-3.11.12-cp312-cp312-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:9dec0000d2d8621d8015c293e24589d46fa218637d820894cb7356c77eca3259"},
-    {file = "aiohttp-3.11.12-cp312-cp312-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:e3552fe98e90fdf5918c04769f338a87fa4f00f3b28830ea9b78b1bdc6140e0d"},
-    {file = "aiohttp-3.11.12-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:6dfe7f984f28a8ae94ff3a7953cd9678550dbd2a1f9bda5dd9c5ae627744c78e"},
-    {file = "aiohttp-3.11.12-cp312-cp312-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:a481a574af914b6e84624412666cbfbe531a05667ca197804ecc19c97b8ab1b0"},
-    {file = "aiohttp-3.11.12-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:1987770fb4887560363b0e1a9b75aa303e447433c41284d3af2840a2f226d6e0"},
-    {file = "aiohttp-3.11.12-cp312-cp312-musllinux_1_2_armv7l.whl", hash = "sha256:a4ac6a0f0f6402854adca4e3259a623f5c82ec3f0c049374133bcb243132baf9"},
-    {file = "aiohttp-3.11.12-cp312-cp312-musllinux_1_2_i686.whl", hash = "sha256:c96a43822f1f9f69cc5c3706af33239489a6294be486a0447fb71380070d4d5f"},
-    {file = "aiohttp-3.11.12-cp312-cp312-musllinux_1_2_ppc64le.whl", hash = "sha256:a5e69046f83c0d3cb8f0d5bd9b8838271b1bc898e01562a04398e160953e8eb9"},
-    {file = "aiohttp-3.11.12-cp312-cp312-musllinux_1_2_s390x.whl", hash = "sha256:68d54234c8d76d8ef74744f9f9fc6324f1508129e23da8883771cdbb5818cbef"},
-    {file = "aiohttp-3.11.12-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:c9fd9dcf9c91affe71654ef77426f5cf8489305e1c66ed4816f5a21874b094b9"},
-    {file = "aiohttp-3.11.12-cp312-cp312-win32.whl", hash = "sha256:0ed49efcd0dc1611378beadbd97beb5d9ca8fe48579fc04a6ed0844072261b6a"},
-    {file = "aiohttp-3.11.12-cp312-cp312-win_amd64.whl", hash = "sha256:54775858c7f2f214476773ce785a19ee81d1294a6bedc5cc17225355aab74802"},
-    {file = "aiohttp-3.11.12-cp313-cp313-macosx_10_13_universal2.whl", hash = "sha256:413ad794dccb19453e2b97c2375f2ca3cdf34dc50d18cc2693bd5aed7d16f4b9"},
-    {file = "aiohttp-3.11.12-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:4a93d28ed4b4b39e6f46fd240896c29b686b75e39cc6992692e3922ff6982b4c"},
-    {file = "aiohttp-3.11.12-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:d589264dbba3b16e8951b6f145d1e6b883094075283dafcab4cdd564a9e353a0"},
-    {file = "aiohttp-3.11.12-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:e5148ca8955affdfeb864aca158ecae11030e952b25b3ae15d4e2b5ba299bad2"},
-    {file = "aiohttp-3.11.12-cp313-cp313-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:525410e0790aab036492eeea913858989c4cb070ff373ec3bc322d700bdf47c1"},
-    {file = "aiohttp-3.11.12-cp313-cp313-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:9bd8695be2c80b665ae3f05cb584093a1e59c35ecb7d794d1edd96e8cc9201d7"},
-    {file = "aiohttp-3.11.12-cp313-cp313-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:f0203433121484b32646a5f5ea93ae86f3d9559d7243f07e8c0eab5ff8e3f70e"},
-    {file = "aiohttp-3.11.12-cp313-cp313-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:40cd36749a1035c34ba8d8aaf221b91ca3d111532e5ccb5fa8c3703ab1b967ed"},
-    {file = "aiohttp-3.11.12-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:a7442662afebbf7b4c6d28cb7aab9e9ce3a5df055fc4116cc7228192ad6cb484"},
-    {file = "aiohttp-3.11.12-cp313-cp313-musllinux_1_2_armv7l.whl", hash = "sha256:8a2fb742ef378284a50766e985804bd6adb5adb5aa781100b09befdbfa757b65"},
-    {file = "aiohttp-3.11.12-cp313-cp313-musllinux_1_2_i686.whl", hash = "sha256:2cee3b117a8d13ab98b38d5b6bdcd040cfb4181068d05ce0c474ec9db5f3c5bb"},
-    {file = "aiohttp-3.11.12-cp313-cp313-musllinux_1_2_ppc64le.whl", hash = "sha256:f6a19bcab7fbd8f8649d6595624856635159a6527861b9cdc3447af288a00c00"},
-    {file = "aiohttp-3.11.12-cp313-cp313-musllinux_1_2_s390x.whl", hash = "sha256:e4cecdb52aaa9994fbed6b81d4568427b6002f0a91c322697a4bfcc2b2363f5a"},
-    {file = "aiohttp-3.11.12-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:30f546358dfa0953db92ba620101fefc81574f87b2346556b90b5f3ef16e55ce"},
-    {file = "aiohttp-3.11.12-cp313-cp313-win32.whl", hash = "sha256:ce1bb21fc7d753b5f8a5d5a4bae99566386b15e716ebdb410154c16c91494d7f"},
-    {file = "aiohttp-3.11.12-cp313-cp313-win_amd64.whl", hash = "sha256:f7914ab70d2ee8ab91c13e5402122edbc77821c66d2758abb53aabe87f013287"},
-    {file = "aiohttp-3.11.12-cp39-cp39-macosx_10_9_universal2.whl", hash = "sha256:7c3623053b85b4296cd3925eeb725e386644fd5bc67250b3bb08b0f144803e7b"},
-    {file = "aiohttp-3.11.12-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:67453e603cea8e85ed566b2700efa1f6916aefbc0c9fcb2e86aaffc08ec38e78"},
-    {file = "aiohttp-3.11.12-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:6130459189e61baac5a88c10019b21e1f0c6d00ebc770e9ce269475650ff7f73"},
-    {file = "aiohttp-3.11.12-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:9060addfa4ff753b09392efe41e6af06ea5dd257829199747b9f15bfad819460"},
-    {file = "aiohttp-3.11.12-cp39-cp39-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:34245498eeb9ae54c687a07ad7f160053911b5745e186afe2d0c0f2898a1ab8a"},
-    {file = "aiohttp-3.11.12-cp39-cp39-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:8dc0fba9a74b471c45ca1a3cb6e6913ebfae416678d90529d188886278e7f3f6"},
-    {file = "aiohttp-3.11.12-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:a478aa11b328983c4444dacb947d4513cb371cd323f3845e53caeda6be5589d5"},
-    {file = "aiohttp-3.11.12-cp39-cp39-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:c160a04283c8c6f55b5bf6d4cad59bb9c5b9c9cd08903841b25f1f7109ef1259"},
-    {file = "aiohttp-3.11.12-cp39-cp39-musllinux_1_2_aarch64.whl", hash = "sha256:edb69b9589324bdc40961cdf0657815df674f1743a8d5ad9ab56a99e4833cfdd"},
-    {file = "aiohttp-3.11.12-cp39-cp39-musllinux_1_2_armv7l.whl", hash = "sha256:4ee84c2a22a809c4f868153b178fe59e71423e1f3d6a8cd416134bb231fbf6d3"},
-    {file = "aiohttp-3.11.12-cp39-cp39-musllinux_1_2_i686.whl", hash = "sha256:bf4480a5438f80e0f1539e15a7eb8b5f97a26fe087e9828e2c0ec2be119a9f72"},
-    {file = "aiohttp-3.11.12-cp39-cp39-musllinux_1_2_ppc64le.whl", hash = "sha256:e6b2732ef3bafc759f653a98881b5b9cdef0716d98f013d376ee8dfd7285abf1"},
-    {file = "aiohttp-3.11.12-cp39-cp39-musllinux_1_2_s390x.whl", hash = "sha256:f752e80606b132140883bb262a457c475d219d7163d996dc9072434ffb0784c4"},
-    {file = "aiohttp-3.11.12-cp39-cp39-musllinux_1_2_x86_64.whl", hash = "sha256:ab3247d58b393bda5b1c8f31c9edece7162fc13265334217785518dd770792b8"},
-    {file = "aiohttp-3.11.12-cp39-cp39-win32.whl", hash = "sha256:0d5176f310a7fe6f65608213cc74f4228e4f4ce9fd10bcb2bb6da8fc66991462"},
-    {file = "aiohttp-3.11.12-cp39-cp39-win_amd64.whl", hash = "sha256:74bd573dde27e58c760d9ca8615c41a57e719bff315c9adb6f2a4281a28e8798"},
-    {file = "aiohttp-3.11.12.tar.gz", hash = "sha256:7603ca26d75b1b86160ce1bbe2787a0b706e592af5b2504e12caa88a217767b0"},
+    {file = "aiohttp-3.11.13-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:a4fe27dbbeec445e6e1291e61d61eb212ee9fed6e47998b27de71d70d3e8777d"},
+    {file = "aiohttp-3.11.13-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:9e64ca2dbea28807f8484c13f684a2f761e69ba2640ec49dacd342763cc265ef"},
+    {file = "aiohttp-3.11.13-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:9840be675de208d1f68f84d578eaa4d1a36eee70b16ae31ab933520c49ba1325"},
+    {file = "aiohttp-3.11.13-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:28a772757c9067e2aee8a6b2b425d0efaa628c264d6416d283694c3d86da7689"},
+    {file = "aiohttp-3.11.13-cp310-cp310-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:b88aca5adbf4625e11118df45acac29616b425833c3be7a05ef63a6a4017bfdb"},
+    {file = "aiohttp-3.11.13-cp310-cp310-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:ce10ddfbe26ed5856d6902162f71b8fe08545380570a885b4ab56aecfdcb07f4"},
+    {file = "aiohttp-3.11.13-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:fa48dac27f41b36735c807d1ab093a8386701bbf00eb6b89a0f69d9fa26b3671"},
+    {file = "aiohttp-3.11.13-cp310-cp310-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:89ce611b1eac93ce2ade68f1470889e0173d606de20c85a012bfa24be96cf867"},
+    {file = "aiohttp-3.11.13-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:78e4dd9c34ec7b8b121854eb5342bac8b02aa03075ae8618b6210a06bbb8a115"},
+    {file = "aiohttp-3.11.13-cp310-cp310-musllinux_1_2_armv7l.whl", hash = "sha256:66047eacbc73e6fe2462b77ce39fc170ab51235caf331e735eae91c95e6a11e4"},
+    {file = "aiohttp-3.11.13-cp310-cp310-musllinux_1_2_i686.whl", hash = "sha256:5ad8f1c19fe277eeb8bc45741c6d60ddd11d705c12a4d8ee17546acff98e0802"},
+    {file = "aiohttp-3.11.13-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:64815c6f02e8506b10113ddbc6b196f58dbef135751cc7c32136df27b736db09"},
+    {file = "aiohttp-3.11.13-cp310-cp310-musllinux_1_2_s390x.whl", hash = "sha256:967b93f21b426f23ca37329230d5bd122f25516ae2f24a9cea95a30023ff8283"},
+    {file = "aiohttp-3.11.13-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:cf1f31f83d16ec344136359001c5e871915c6ab685a3d8dee38e2961b4c81730"},
+    {file = "aiohttp-3.11.13-cp310-cp310-win32.whl", hash = "sha256:00c8ac69e259c60976aa2edae3f13d9991cf079aaa4d3cd5a49168ae3748dee3"},
+    {file = "aiohttp-3.11.13-cp310-cp310-win_amd64.whl", hash = "sha256:90d571c98d19a8b6e793b34aa4df4cee1e8fe2862d65cc49185a3a3d0a1a3996"},
+    {file = "aiohttp-3.11.13-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:6b35aab22419ba45f8fc290d0010898de7a6ad131e468ffa3922b1b0b24e9d2e"},
+    {file = "aiohttp-3.11.13-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:f81cba651db8795f688c589dd11a4fbb834f2e59bbf9bb50908be36e416dc760"},
+    {file = "aiohttp-3.11.13-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:f55d0f242c2d1fcdf802c8fabcff25a9d85550a4cf3a9cf5f2a6b5742c992839"},
+    {file = "aiohttp-3.11.13-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:c4bea08a6aad9195ac9b1be6b0c7e8a702a9cec57ce6b713698b4a5afa9c2e33"},
+    {file = "aiohttp-3.11.13-cp311-cp311-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:c6070bcf2173a7146bb9e4735b3c62b2accba459a6eae44deea0eb23e0035a23"},
+    {file = "aiohttp-3.11.13-cp311-cp311-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:718d5deb678bc4b9d575bfe83a59270861417da071ab44542d0fcb6faa686636"},
+    {file = "aiohttp-3.11.13-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:0f6b2c5b4a4d22b8fb2c92ac98e0747f5f195e8e9448bfb7404cd77e7bfa243f"},
+    {file = "aiohttp-3.11.13-cp311-cp311-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:747ec46290107a490d21fe1ff4183bef8022b848cf9516970cb31de6d9460088"},
+    {file = "aiohttp-3.11.13-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:01816f07c9cc9d80f858615b1365f8319d6a5fd079cd668cc58e15aafbc76a54"},
+    {file = "aiohttp-3.11.13-cp311-cp311-musllinux_1_2_armv7l.whl", hash = "sha256:a08ad95fcbd595803e0c4280671d808eb170a64ca3f2980dd38e7a72ed8d1fea"},
+    {file = "aiohttp-3.11.13-cp311-cp311-musllinux_1_2_i686.whl", hash = "sha256:c97be90d70f7db3aa041d720bfb95f4869d6063fcdf2bb8333764d97e319b7d0"},
+    {file = "aiohttp-3.11.13-cp311-cp311-musllinux_1_2_ppc64le.whl", hash = "sha256:ab915a57c65f7a29353c8014ac4be685c8e4a19e792a79fe133a8e101111438e"},
+    {file = "aiohttp-3.11.13-cp311-cp311-musllinux_1_2_s390x.whl", hash = "sha256:35cda4e07f5e058a723436c4d2b7ba2124ab4e0aa49e6325aed5896507a8a42e"},
+    {file = "aiohttp-3.11.13-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:af55314407714fe77a68a9ccaab90fdb5deb57342585fd4a3a8102b6d4370080"},
+    {file = "aiohttp-3.11.13-cp311-cp311-win32.whl", hash = "sha256:42d689a5c0a0c357018993e471893e939f555e302313d5c61dfc566c2cad6185"},
+    {file = "aiohttp-3.11.13-cp311-cp311-win_amd64.whl", hash = "sha256:b73a2b139782a07658fbf170fe4bcdf70fc597fae5ffe75e5b67674c27434a9f"},
+    {file = "aiohttp-3.11.13-cp312-cp312-macosx_10_13_universal2.whl", hash = "sha256:2eabb269dc3852537d57589b36d7f7362e57d1ece308842ef44d9830d2dc3c90"},
+    {file = "aiohttp-3.11.13-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:7b77ee42addbb1c36d35aca55e8cc6d0958f8419e458bb70888d8c69a4ca833d"},
+    {file = "aiohttp-3.11.13-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:55789e93c5ed71832e7fac868167276beadf9877b85697020c46e9a75471f55f"},
+    {file = "aiohttp-3.11.13-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:c929f9a7249a11e4aa5c157091cfad7f49cc6b13f4eecf9b747104befd9f56f2"},
+    {file = "aiohttp-3.11.13-cp312-cp312-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:d33851d85537bbf0f6291ddc97926a754c8f041af759e0aa0230fe939168852b"},
+    {file = "aiohttp-3.11.13-cp312-cp312-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:9229d8613bd8401182868fe95688f7581673e1c18ff78855671a4b8284f47bcb"},
+    {file = "aiohttp-3.11.13-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:669dd33f028e54fe4c96576f406ebb242ba534dd3a981ce009961bf49960f117"},
+    {file = "aiohttp-3.11.13-cp312-cp312-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:7c1b20a1ace54af7db1f95af85da530fe97407d9063b7aaf9ce6a32f44730778"},
+    {file = "aiohttp-3.11.13-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:5724cc77f4e648362ebbb49bdecb9e2b86d9b172c68a295263fa072e679ee69d"},
+    {file = "aiohttp-3.11.13-cp312-cp312-musllinux_1_2_armv7l.whl", hash = "sha256:aa36c35e94ecdb478246dd60db12aba57cfcd0abcad43c927a8876f25734d496"},
+    {file = "aiohttp-3.11.13-cp312-cp312-musllinux_1_2_i686.whl", hash = "sha256:9b5b37c863ad5b0892cc7a4ceb1e435e5e6acd3f2f8d3e11fa56f08d3c67b820"},
+    {file = "aiohttp-3.11.13-cp312-cp312-musllinux_1_2_ppc64le.whl", hash = "sha256:e06cf4852ce8c4442a59bae5a3ea01162b8fcb49ab438d8548b8dc79375dad8a"},
+    {file = "aiohttp-3.11.13-cp312-cp312-musllinux_1_2_s390x.whl", hash = "sha256:5194143927e494616e335d074e77a5dac7cd353a04755330c9adc984ac5a628e"},
+    {file = "aiohttp-3.11.13-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:afcb6b275c2d2ba5d8418bf30a9654fa978b4f819c2e8db6311b3525c86fe637"},
+    {file = "aiohttp-3.11.13-cp312-cp312-win32.whl", hash = "sha256:7104d5b3943c6351d1ad7027d90bdd0ea002903e9f610735ac99df3b81f102ee"},
+    {file = "aiohttp-3.11.13-cp312-cp312-win_amd64.whl", hash = "sha256:47dc018b1b220c48089b5b9382fbab94db35bef2fa192995be22cbad3c5730c8"},
+    {file = "aiohttp-3.11.13-cp313-cp313-macosx_10_13_universal2.whl", hash = "sha256:9862d077b9ffa015dbe3ce6c081bdf35135948cb89116e26667dd183550833d1"},
+    {file = "aiohttp-3.11.13-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:fbfef0666ae9e07abfa2c54c212ac18a1f63e13e0760a769f70b5717742f3ece"},
+    {file = "aiohttp-3.11.13-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:93a1f7d857c4fcf7cabb1178058182c789b30d85de379e04f64c15b7e88d66fb"},
+    {file = "aiohttp-3.11.13-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:ba40b7ae0f81c7029583a338853f6607b6d83a341a3dcde8bed1ea58a3af1df9"},
+    {file = "aiohttp-3.11.13-cp313-cp313-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:b5b95787335c483cd5f29577f42bbe027a412c5431f2f80a749c80d040f7ca9f"},
+    {file = "aiohttp-3.11.13-cp313-cp313-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:a7d474c5c1f0b9405c1565fafdc4429fa7d986ccbec7ce55bc6a330f36409cad"},
+    {file = "aiohttp-3.11.13-cp313-cp313-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:1e83fb1991e9d8982b3b36aea1e7ad27ea0ce18c14d054c7a404d68b0319eebb"},
+    {file = "aiohttp-3.11.13-cp313-cp313-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:4586a68730bd2f2b04a83e83f79d271d8ed13763f64b75920f18a3a677b9a7f0"},
+    {file = "aiohttp-3.11.13-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:9fe4eb0e7f50cdb99b26250d9328faef30b1175a5dbcfd6d0578d18456bac567"},
+    {file = "aiohttp-3.11.13-cp313-cp313-musllinux_1_2_armv7l.whl", hash = "sha256:2a8a6bc19818ac3e5596310ace5aa50d918e1ebdcc204dc96e2f4d505d51740c"},
+    {file = "aiohttp-3.11.13-cp313-cp313-musllinux_1_2_i686.whl", hash = "sha256:7f27eec42f6c3c1df09cfc1f6786308f8b525b8efaaf6d6bd76c1f52c6511f6a"},
+    {file = "aiohttp-3.11.13-cp313-cp313-musllinux_1_2_ppc64le.whl", hash = "sha256:2a4a13dfbb23977a51853b419141cd0a9b9573ab8d3a1455c6e63561387b52ff"},
+    {file = "aiohttp-3.11.13-cp313-cp313-musllinux_1_2_s390x.whl", hash = "sha256:02876bf2f69b062584965507b07bc06903c2dc93c57a554b64e012d636952654"},
+    {file = "aiohttp-3.11.13-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:b992778d95b60a21c4d8d4a5f15aaab2bd3c3e16466a72d7f9bfd86e8cea0d4b"},
+    {file = "aiohttp-3.11.13-cp313-cp313-win32.whl", hash = "sha256:507ab05d90586dacb4f26a001c3abf912eb719d05635cbfad930bdbeb469b36c"},
+    {file = "aiohttp-3.11.13-cp313-cp313-win_amd64.whl", hash = "sha256:5ceb81a4db2decdfa087381b5fc5847aa448244f973e5da232610304e199e7b2"},
+    {file = "aiohttp-3.11.13-cp39-cp39-macosx_10_9_universal2.whl", hash = "sha256:51c3ff9c7a25f3cad5c09d9aacbc5aefb9267167c4652c1eb737989b554fe278"},
+    {file = "aiohttp-3.11.13-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:e271beb2b1dabec5cd84eb488bdabf9758d22ad13471e9c356be07ad139b3012"},
+    {file = "aiohttp-3.11.13-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:0e9eb7e5764abcb49f0e2bd8f5731849b8728efbf26d0cac8e81384c95acec3f"},
+    {file = "aiohttp-3.11.13-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:baae005092e3f200de02699314ac8933ec20abf998ec0be39448f6605bce93df"},
+    {file = "aiohttp-3.11.13-cp39-cp39-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:1982c98ac62c132d2b773d50e2fcc941eb0b8bad3ec078ce7e7877c4d5a2dce7"},
+    {file = "aiohttp-3.11.13-cp39-cp39-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:d2b25b2eeb35707113b2d570cadc7c612a57f1c5d3e7bb2b13870fe284e08fc0"},
+    {file = "aiohttp-3.11.13-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:b27961d65639128336b7a7c3f0046dcc62a9443d5ef962e3c84170ac620cec47"},
+    {file = "aiohttp-3.11.13-cp39-cp39-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:a01fe9f1e05025eacdd97590895e2737b9f851d0eb2e017ae9574d9a4f0b6252"},
+    {file = "aiohttp-3.11.13-cp39-cp39-musllinux_1_2_aarch64.whl", hash = "sha256:fa1fb1b61881c8405829c50e9cc5c875bfdbf685edf57a76817dfb50643e4a1a"},
+    {file = "aiohttp-3.11.13-cp39-cp39-musllinux_1_2_armv7l.whl", hash = "sha256:25de43bb3cf83ad83efc8295af7310219af6dbe4c543c2e74988d8e9c8a2a917"},
+    {file = "aiohttp-3.11.13-cp39-cp39-musllinux_1_2_i686.whl", hash = "sha256:fe7065e2215e4bba63dc00db9ae654c1ba3950a5fff691475a32f511142fcddb"},
+    {file = "aiohttp-3.11.13-cp39-cp39-musllinux_1_2_ppc64le.whl", hash = "sha256:7836587eef675a17d835ec3d98a8c9acdbeb2c1d72b0556f0edf4e855a25e9c1"},
+    {file = "aiohttp-3.11.13-cp39-cp39-musllinux_1_2_s390x.whl", hash = "sha256:85fa0b18558eb1427090912bd456a01f71edab0872f4e0f9e4285571941e4090"},
+    {file = "aiohttp-3.11.13-cp39-cp39-musllinux_1_2_x86_64.whl", hash = "sha256:a86dc177eb4c286c19d1823ac296299f59ed8106c9536d2b559f65836e0fb2c6"},
+    {file = "aiohttp-3.11.13-cp39-cp39-win32.whl", hash = "sha256:684eea71ab6e8ade86b9021bb62af4bf0881f6be4e926b6b5455de74e420783a"},
+    {file = "aiohttp-3.11.13-cp39-cp39-win_amd64.whl", hash = "sha256:82c249f2bfa5ecbe4a1a7902c81c0fba52ed9ebd0176ab3047395d02ad96cfcb"},
+    {file = "aiohttp-3.11.13.tar.gz", hash = "sha256:8ce789231404ca8fff7f693cdce398abf6d90fd5dae2b1847477196c243b1fbb"},
 ]
 
 [package.dependencies]
@@ -959,13 +959,13 @@ files = [
 
 [[package]]
 name = "decorator"
-version = "5.1.1"
+version = "5.2.1"
 description = "Decorators for Humans"
 optional = false
-python-versions = ">=3.5"
+python-versions = ">=3.8"
 files = [
-    {file = "decorator-5.1.1-py3-none-any.whl", hash = "sha256:b8c3f85900b9dc423225913c5aace94729fe1fa9763b38939a95226f02d37186"},
-    {file = "decorator-5.1.1.tar.gz", hash = "sha256:637996211036b6385ef91435e4fae22989472f9d571faba8927ba8253acbc330"},
+    {file = "decorator-5.2.1-py3-none-any.whl", hash = "sha256:d316bb415a2d9e2d2b3abcc4084c6502fc09240e292cd76a76afc106a1c8e04a"},
+    {file = "decorator-5.2.1.tar.gz", hash = "sha256:65f266143752f734b0a7cc83c46f4618af75b8c5911b00ccb61d0ac9b6da0360"},
 ]
 
 [[package]]
@@ -2707,21 +2707,21 @@ dill = ">=0.3.8"
 
 [[package]]
 name = "narwhals"
-version = "1.27.1"
+version = "1.28.0"
 description = "Extremely lightweight compatibility layer between dataframe libraries"
 optional = false
 python-versions = ">=3.8"
 files = [
-    {file = "narwhals-1.27.1-py3-none-any.whl", hash = "sha256:71e4a126007886e3dd9d71d0d5921ebd2e8c1f9be9c405fe11850ece2b066c59"},
-    {file = "narwhals-1.27.1.tar.gz", hash = "sha256:68505d0cee1e6c00382ac8b65e922f8b694a11cbe482a057fa63139de8d0ea03"},
+    {file = "narwhals-1.28.0-py3-none-any.whl", hash = "sha256:45d909ad6240944d447b0dae38074c5a919830dff3868d57b05a5526c1f06fe4"},
+    {file = "narwhals-1.28.0.tar.gz", hash = "sha256:a2213fa44a039f724278fb15609889319e7c240403413f2606cc856c8d8f708d"},
 ]
 
 [package.extras]
 core = ["duckdb", "pandas", "polars", "pyarrow", "pyarrow-stubs"]
 cudf = ["cudf (>=24.10.0)"]
 dask = ["dask[dataframe] (>=2024.8)"]
-dev = ["covdefaults", "hypothesis", "mypy (>=1.15.0,<1.16.0)", "pandas-stubs", "pre-commit", "pytest", "pytest-cov", "pytest-env", "pytest-randomly", "typing-extensions"]
-docs = ["black", "duckdb", "jinja2", "markdown-exec[ansi]", "mkdocs", "mkdocs-autorefs", "mkdocs-material", "mkdocstrings[python]", "pandas", "polars (>=1.0.0)", "pyarrow"]
+dev = ["covdefaults", "hypothesis", "mypy (>=1.15.0,<1.16.0)", "pandas-stubs", "pre-commit", "pyright", "pytest", "pytest-cov", "pytest-env", "pytest-randomly", "typing-extensions"]
+docs = ["black", "duckdb", "jinja2", "markdown-exec[ansi]", "mkdocs", "mkdocs-autorefs", "mkdocs-material", "mkdocstrings-python (>=1.16)", "mkdocstrings[python]", "pandas", "polars (>=1.0.0)", "pyarrow"]
 duckdb = ["duckdb (>=1.0)"]
 extra = ["scikit-learn"]
 ibis = ["ibis-framework (>=6.0.0)", "packaging", "pyarrow-hotfix", "rich"]
@@ -2731,7 +2731,7 @@ polars = ["polars (>=0.20.3)"]
 pyarrow = ["pyarrow (>=11.0.0)"]
 pyspark = ["pyspark (>=3.5.0)"]
 tests = ["covdefaults", "hypothesis", "pytest", "pytest-cov", "pytest-env", "pytest-randomly", "typing-extensions"]
-typing = ["mypy (>=1.15.0,<1.16.0)", "pandas-stubs", "typing-extensions"]
+typing = ["mypy (>=1.15.0,<1.16.0)", "pandas-stubs", "pyright", "typing-extensions"]
 
 [[package]]
 name = "nbclient"
@@ -5074,13 +5074,13 @@ snowflake = ["snowflake-connector-python (>=3.3.0)", "snowflake-snowpark-python[
 
 [[package]]
 name = "surya-ocr"
-version = "0.12.0"
+version = "0.12.1"
 description = "OCR, layout, reading order, and table recognition in 90+ languages"
 optional = false
 python-versions = "<4.0,>=3.10"
 files = [
-    {file = "surya_ocr-0.12.0-py3-none-any.whl", hash = "sha256:423e591edc97212743d8b013f5cf123ba4c461d44858edf1b67323fbede9c9b5"},
-    {file = "surya_ocr-0.12.0.tar.gz", hash = "sha256:d134d8d7590bd034042e344f1d04ebf8cd96f2f504d5a38eed72caa4196b4081"},
+    {file = "surya_ocr-0.12.1-py3-none-any.whl", hash = "sha256:703362d808994576e7cd297c731c74af552c1be831cbb933968736a187d89ac3"},
+    {file = "surya_ocr-0.12.1.tar.gz", hash = "sha256:4b37d94db3747f843c23c9da4f146c436de514945297bdebfc394971c03f6340"},
 ]
 
 [package.dependencies]
@@ -6107,4 +6107,4 @@ test = ["pytest"]
 [metadata]
 lock-version = "2.0"
 python-versions = "^3.10"
-content-hash = "307a7d77aa28ba7ae943396fbc36e03cd1b0c4dd354ed9e7cfbf7cd74c49e164"
+content-hash = "0a6d5e377e87278aa87ec01a22dfcb38fa9e9083136fe73d3a9d1bec02683dae"
diff --git a/pyproject.toml b/pyproject.toml
index 379e4e04..d23f78b5 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -26,7 +26,7 @@ torch = "^2.5.1"
 tqdm = "^4.66.1"
 ftfy = "^6.1.1"
 rapidfuzz = "^3.8.1"
-surya-ocr = "~0.12.0"
+surya-ocr = "~0.12.1"
 regex = "^2024.4.28"
 pdftext = "~0.6.0"
 markdownify = "^0.13.1"

From 6d7d95ece5576cd6ff00f74c13842a344268356e Mon Sep 17 00:00:00 2001
From: Vik Paruchuri <vik.paruchuri@gmail.com>
Date: Mon, 24 Feb 2025 11:57:05 -0500
Subject: [PATCH 20/46] Bump README

---
 README.md                 | 2 +-
 marker/services/vertex.py | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/README.md b/README.md
index e3cc5f19..9a65f875 100644
--- a/README.md
+++ b/README.md
@@ -2,7 +2,7 @@
 
 Marker converts documents to markdown, JSON, and HTML quickly and accurately.
 
-- Converts PDF, image, PPTX, DOCX, XLSX, HTML, EPUB in all languages
+- Converts PDF, image, PPTX, DOCX, XLSX, HTML, EPUB files in all languages
 - Formats tables, forms, equations, inline math, links, references, and code blocks
 - Extracts and saves images
 - Removes headers/footers/other artifacts
diff --git a/marker/services/vertex.py b/marker/services/vertex.py
index 35dfb946..a773264d 100644
--- a/marker/services/vertex.py
+++ b/marker/services/vertex.py
@@ -16,7 +16,7 @@ class GoogleVertexService(BaseGeminiService):
     gemini_model_name: Annotated[
         str,
         "The name of the Google model to use for the service."
-    ] = "gemini-1.5-flash-002"
+    ] = "gemini-2.0-flash-001"
 
     def get_google_client(self, timeout: int):
         return genai.Client(

From 9d91954c9e02f1b9366c4df7b5edc85bce2b0cd1 Mon Sep 17 00:00:00 2001
From: Vik Paruchuri <vik.paruchuri@gmail.com>
Date: Mon, 24 Feb 2025 12:14:01 -0500
Subject: [PATCH 21/46] Test fixes

---
 tests/builders/test_inline_math_lines.py | 2 +-
 tests/processors/test_inline_math.py     | 4 ++--
 2 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/tests/builders/test_inline_math_lines.py b/tests/builders/test_inline_math_lines.py
index a718c3b6..faa3f411 100644
--- a/tests/builders/test_inline_math_lines.py
+++ b/tests/builders/test_inline_math_lines.py
@@ -14,7 +14,7 @@ def test_inline_box_nomerging(pdf_document, config):
     merger(pdf_document)
 
     line_count = len(block.contained_blocks(pdf_document, (BlockTypes.Line,)))
-    assert line_count == 45
+    assert line_count == 46
 
 
 @pytest.mark.config({"page_range": [1], "use_llm": True})
diff --git a/tests/processors/test_inline_math.py b/tests/processors/test_inline_math.py
index 4abb6972..a609ba38 100644
--- a/tests/processors/test_inline_math.py
+++ b/tests/processors/test_inline_math.py
@@ -13,7 +13,7 @@ def test_llm_text_processor(pdf_document, mocker):
     # Get all inline math lines
     text_lines = pdf_document.contained_blocks((BlockTypes.Line,))
     text_lines = [line for line in text_lines if line.formats and "math" in line.formats]
-    assert len(text_lines) == 3
+    assert len(text_lines) == 8
     corrected_lines = ["<math>Text</math>"] * len(text_lines)
 
     mock_cls = Mock()
@@ -44,4 +44,4 @@ def test_llm_text_processor_texify(pdf_document):
     # Get all inline math lines
     text_lines = pdf_document.contained_blocks((BlockTypes.Line,))
     text_lines = [line for line in text_lines if line.formats and "math" in line.formats]
-    assert len(text_lines) == 3
\ No newline at end of file
+    assert len(text_lines) == 8
\ No newline at end of file

From 6e0afa23644dca180c2e5876fc99198b1dc06372 Mon Sep 17 00:00:00 2001
From: Vik Paruchuri <vik.paruchuri@gmail.com>
Date: Mon, 24 Feb 2025 16:47:46 -0500
Subject: [PATCH 22/46] Bump version

---
 marker/processors/debug.py                     | 3 +++
 marker/processors/llm/llm_image_description.py | 2 +-
 pyproject.toml                                 | 4 ++--
 3 files changed, 6 insertions(+), 3 deletions(-)

diff --git a/marker/processors/debug.py b/marker/processors/debug.py
index cf9f0da6..e6c53b82 100644
--- a/marker/processors/debug.py
+++ b/marker/processors/debug.py
@@ -93,6 +93,9 @@ def draw_layout_debug_images(self, document: Document, pdf_mode=False):
             line_bboxes = []
             line_text = []
             for child in page.children:
+                if child.removed:
+                    continue
+
                 if child.block_type != BlockTypes.Line:
                     continue
 
diff --git a/marker/processors/llm/llm_image_description.py b/marker/processors/llm/llm_image_description.py
index 7b768ef6..83eeacc2 100644
--- a/marker/processors/llm/llm_image_description.py
+++ b/marker/processors/llm/llm_image_description.py
@@ -23,7 +23,7 @@ class LLMImageDescriptionProcessor(BaseLLMSimpleBlockProcessor):
 **Instructions:**
 1. Carefully examine the provided image.
 2. Analyze any text that was extracted from within the image.
-3. Output a 3-4 sentence description of the image.  Make sure there is enough specific detail to accurately describe the image.  If there are numbers included, try to be specific.
+3. Output a faithful description of the image.  Make sure there is enough specific detail to accurately reconstruct the image.  If the image is a figure or contains numeric data, include the numeric data in the output.
 **Example:**
 Input:
 ```text
diff --git a/pyproject.toml b/pyproject.toml
index d23f78b5..32353d92 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -1,7 +1,7 @@
 [tool.poetry]
 name = "marker-pdf"
-version = "1.5.6"
-description = "Convert PDF to markdown with high speed and accuracy."
+version = "1.6.0"
+description = "Convert documents to markdown with high speed and accuracy."
 authors = ["Vik Paruchuri <github@vikas.sh>"]
 readme = "README.md"
 license = "GPL-3.0-or-later"

From 118c12b175eb20e92b4874d76d5f87dfea529a4e Mon Sep 17 00:00:00 2001
From: Vik Paruchuri <vik.paruchuri@gmail.com>
Date: Mon, 24 Feb 2025 16:59:12 -0500
Subject: [PATCH 23/46] Make merges simpler

---
 marker/processors/line_merge.py | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/marker/processors/line_merge.py b/marker/processors/line_merge.py
index 9c45bde9..f8045c2f 100644
--- a/marker/processors/line_merge.py
+++ b/marker/processors/line_merge.py
@@ -28,7 +28,7 @@ class LineMergeProcessor(BaseProcessor):
     intersection_pct_threshold: Annotated[
         float,
         "The total amount of intersection area concentrated in the max intersection block."
-    ] = .6
+    ] = .5
     vertical_overlap_pct_threshold: Annotated[
         float,
         "The minimum percentage of vertical overlap to consider merging."
@@ -80,7 +80,7 @@ def merge_lines(self, lines: List[Line], block: Block):
                 # Within same line
                 vertical_overlap_pct > self.vertical_overlap_pct_threshold,
                 # doesn't overlap with anything else
-                merge_intersection / total_intersection >= self.intersection_pct_threshold
+                merge_intersection / total_intersection > self.intersection_pct_threshold
             ]):
                 merge.append(i)
             else:

From c2dd7e14f491c63685130431a11262721c0c3b8b Mon Sep 17 00:00:00 2001
From: Vik Paruchuri <vik.paruchuri@gmail.com>
Date: Mon, 24 Feb 2025 18:32:07 -0500
Subject: [PATCH 24/46] Adjust merging

---
 marker/builders/line.py      | 93 ++++++++++++++++++++----------------
 marker/schema/blocks/base.py |  2 +
 marker/services/__init__.py  |  2 +-
 3 files changed, 55 insertions(+), 42 deletions(-)

diff --git a/marker/builders/line.py b/marker/builders/line.py
index 4133c374..752cd28a 100644
--- a/marker/builders/line.py
+++ b/marker/builders/line.py
@@ -421,55 +421,66 @@ def merge_provider_lines_inline_math(
             merge_lines[best_overlap].append(i)
 
         # Filter to get rid of detected lines that include multiple provider lines
-        filtered_merge_lines = {}
+        filtered_merge_lines = defaultdict(list)
         for line_idx in merge_lines:
-            first_line = horizontal_provider_lines[merge_lines[line_idx][0]][1].line.polygon
-            all_close = all([
-                (
-                    abs(horizontal_provider_lines[ml][1].line.polygon.y_start - first_line.y_start) < self.inline_math_line_vertical_merge_threshold
-                    or
-                    abs(horizontal_provider_lines[ml][1].line.polygon.y_end - first_line.y_end) < self.inline_math_line_vertical_merge_threshold
-                )
-                for ml in
-                merge_lines[line_idx]
-            ])
-            if all_close:
-                filtered_merge_lines[line_idx] = merge_lines[line_idx]
+            merge_segment = []
+            prev_line = None
+            for ml in merge_lines[line_idx]:
+                line = horizontal_provider_lines[ml][1].line.polygon
+                if prev_line:
+                    close = (
+                        abs(line.y_start - prev_line.y_start) < self.inline_math_line_vertical_merge_threshold
+                        or
+                        abs(line.y_end - prev_line.y_end) < self.inline_math_line_vertical_merge_threshold
+                    )
+                else:
+                    # First line
+                    close = True
+
+                prev_line = line
+                if close:
+                    merge_segment.append(ml)
+                else:
+                    if merge_segment:
+                        filtered_merge_lines[line_idx].append(merge_segment)
+                    merge_segment = [ml]
 
         # Handle the merging
         already_merged = set()
-        potential_merges = set(chain.from_iterable(filtered_merge_lines.values()))
+        potential_merges = []
+        for line_idx in filtered_merge_lines:
+            potential_merges.extend(chain.from_iterable(filtered_merge_lines[line_idx]))
+        potential_merges = set(potential_merges)
         out_provider_lines = [(i, p) for i, p in enumerate(provider_lines) if i not in potential_merges]
         for line_idx in filtered_merge_lines:
             text_line = text_lines[line_idx]
-            merge_section = filtered_merge_lines[line_idx]
-            merge_section = [m for m in merge_section if m not in already_merged]
-            if len(merge_section) == 0:
-                continue
-            elif len(merge_section) == 1:
-                line_idx = merge_section[0]
-                merged_line = provider_lines[line_idx]
-                # Only add math format to single lines if the detected line is math
-                if text_line.math:
-                    self.add_math_span_format(merged_line)
-                out_provider_lines.append((line_idx, merged_line))
-                already_merged.add(merge_section[0])
-                continue
-
-            merge_section = sorted(merge_section)
-            merged_line = None
-            min_idx = min(merge_section)
-            for idx in merge_section:
-                provider_line = deepcopy(provider_lines[idx])
-                if merged_line is None:
-                    merged_line = provider_line
+            for merge_section in filtered_merge_lines[line_idx]:
+                merge_section = [m for m in merge_section if m not in already_merged]
+                if len(merge_section) == 0:
+                    continue
+                elif len(merge_section) == 1:
+                    provider_idx = merge_section[0]
+                    merged_line = provider_lines[provider_idx]
+                    # Only add math format to single lines if the detected line is math
+                    if text_line.math:
+                        self.add_math_span_format(merged_line)
+                    out_provider_lines.append((provider_idx, merged_line))
+                    already_merged.add(merge_section[0])
                 else:
-                    # Combine the spans of the provider line with the merged line
-                    merged_line = merged_line.merge(provider_line)
-                    # Add math regardless, since we assume heavily broken lines are math lines
-                    self.add_math_span_format(merged_line)
-                already_merged.add(idx) # Prevent double merging
-            out_provider_lines.append((min_idx, merged_line))
+                    merge_section = sorted(merge_section)
+                    merged_line = None
+                    min_idx = min(merge_section)
+                    for idx in merge_section:
+                        provider_line = deepcopy(provider_lines[idx])
+                        if merged_line is None:
+                            merged_line = provider_line
+                        else:
+                            # Combine the spans of the provider line with the merged line
+                            merged_line = merged_line.merge(provider_line)
+                            # Add math regardless, since we assume heavily broken lines are math lines
+                            self.add_math_span_format(merged_line)
+                        already_merged.add(idx) # Prevent double merging
+                    out_provider_lines.append((min_idx, merged_line))
 
         # Sort to preserve original order
         out_provider_lines = sorted(out_provider_lines, key=lambda x: x[0])
diff --git a/marker/schema/blocks/base.py b/marker/schema/blocks/base.py
index f68ea799..6bfaa9fc 100644
--- a/marker/schema/blocks/base.py
+++ b/marker/schema/blocks/base.py
@@ -215,6 +215,8 @@ def contained_blocks(self, document: Document, block_types: Sequence[BlockTypes]
         blocks = []
         for block_id in self.structure:
             block = document.get_block(block_id)
+            if block.removed:
+                continue
             if (block_types is None or block.block_type in block_types) and not block.removed:
                 blocks.append(block)
             blocks += block.contained_blocks(document, block_types)
diff --git a/marker/services/__init__.py b/marker/services/__init__.py
index 7f9d3a4b..0c7fe1c0 100644
--- a/marker/services/__init__.py
+++ b/marker/services/__init__.py
@@ -15,7 +15,7 @@ class BaseService:
     max_retries: Annotated[
         int,
         "The maximum number of retries to use for the service."
-    ] = 1
+    ] = 3
 
     def __init__(self, config: Optional[BaseModel | dict] = None):
         assign_config(self, config)

From 0c8fc1f7e26bffeefb0eb9ccd1b99133f143e002 Mon Sep 17 00:00:00 2001
From: Vik Paruchuri <vik.paruchuri@gmail.com>
Date: Mon, 24 Feb 2025 21:45:27 -0500
Subject: [PATCH 25/46] Improve streamlit app

---
 marker/scripts/streamlit_app.py              |   2 +-
 marker/scripts/streamlit_app_blocks_viz.html | 130 ++++++++++++++-----
 2 files changed, 98 insertions(+), 34 deletions(-)

diff --git a/marker/scripts/streamlit_app.py b/marker/scripts/streamlit_app.py
index 96d97234..4bc8acaf 100644
--- a/marker/scripts/streamlit_app.py
+++ b/marker/scripts/streamlit_app.py
@@ -134,7 +134,7 @@ def block_display(image: Image, blocks: dict | None = None, dpi=96):
     }
     return components.html(
         BLOCKS_VIZ_TMPL.substitute(**template_values),
-        height=image.height, width=image.width
+        height=image.height
     )
 
 
diff --git a/marker/scripts/streamlit_app_blocks_viz.html b/marker/scripts/streamlit_app_blocks_viz.html
index e00908d1..464803a9 100644
--- a/marker/scripts/streamlit_app_blocks_viz.html
+++ b/marker/scripts/streamlit_app_blocks_viz.html
@@ -7,21 +7,26 @@
         body {
             font-family: "Source Sans Pro",sans-serif;
             font-weight: 400;
-            -moz-osx-font-smoothing: auto
+            -moz-osx-font-smoothing: auto;
+            margin: 0;
+            padding: 0;
         }
 
         .tippy-box {
-            font-size: 10px
+            font-size: clamp(8px, 2vw, 10px);
         }
 
         .image-container {
             position: relative;
-            width: 90%
+            width: 100%;
+            max-width: 1200px;
+            margin: 0 auto;
         }
 
         .image-container img {
             width: 100%;
-            height: auto
+            height: auto;
+            display: block;
         }
 
         .blocks-overlay {
@@ -29,72 +34,90 @@
             top: 0;
             left: 0;
             width: 100%;
-            height: 100%
+            height: 100%;
+            pointer-events: none; /* This ensures clicks go through to SVG elements */
         }
 
         .blocks-overlay rect.block {
             fill-opacity: .2;
-            stroke-opacity: .5
+            stroke-opacity: .5;
+            pointer-events: auto; /* This ensures clicks are captured by rectangles */
         }
 
         .blocks-overlay rect.block:hover {
             stroke-opacity: 1;
-            cursor: pointer
+            cursor: pointer;
         }
 
         #block-info-dialog {
-            width: 65%
+            width: 80%;
+            max-width: 600px;
+            max-height: 80vh;
+            border-radius: 8px;
+            border: 1px solid #ccc;
+            padding: 16px;
+            box-shadow: 0 4px 8px rgba(0,0,0,0.2);
         }
 
         #block-info-dialog button.close-button {
-            font-size: 20px;
+            font-size: clamp(16px, 3vw, 20px);
             position: absolute;
-            top: 0;
-            right: 0;
+            top: 8px;
+            right: 8px;
             margin: 0;
             border: 0;
             background: 0 0;
-            padding: 0 4px 0 0;
-            cursor: pointer
+            padding: 0 4px;
+            cursor: pointer;
+            width: 24px;
+            height: 24px;
+            display: flex;
+            align-items: center;
+            justify-content: center;
         }
 
         #block-info-dialog button.close-button:focus {
-            outline: 0
+            outline: 0;
         }
 
         #block-info-dialog button.close-button::after {
-            content: "╳"
+            content: "╳";
         }
 
         #block-info-dialog button.copy-json-button {
-            font-size: 10px;
+            font-size: clamp(8px, 2vw, 10px);
             color: #bababa;
             cursor: pointer;
             position: absolute;
-            bottom: 3px;
-            right: 3px;
+            bottom: 8px;
+            right: 8px;
             border: 0;
-            background: 0 0
+            background: 0 0;
+            padding: 4px 8px;
         }
 
         #block-info-dialog button.copy-json-button:hover {
-            color: #666
+            color: #666;
         }
 
         #block-info-dialog button.copy-json-button:active {
-            color: #000
+            color: #000;
         }
 
         #block-info-dialog h1 {
-            margin: 0 0 10px;
+            margin: 0 0 16px;
             text-align: left;
-            font-size: 1em
+            font-size: clamp(14px, 3vw, 18px);
+            padding-right: 24px; /* Space for close button */
         }
 
         #block-info-dialog .text-content {
             overflow-y: auto;
             font-family: monospace;
-            white-space: pre
+            white-space: pre-wrap;
+            font-size: clamp(10px, 2vw, 14px);
+            max-height: 40vh;
+            margin-bottom: 16px;
         }
 
         #block-info-dialog .images {
@@ -102,17 +125,34 @@
             flex-wrap: wrap;
             justify-content: center;
             gap: 10px;
-            margin-top: 10px
+            margin-top: 10px;
+            margin-bottom: 24px; /* Space for copy button */
         }
 
         #block-info-dialog .images img {
-            max-width: 40%;
-            height: auto
+            max-width: 45%;
+            height: auto;
+        }
+
+        @media screen and (max-width: 768px) {
+            #block-info-dialog {
+                width: 90%;
+            }
+
+            #block-info-dialog .images img {
+                max-width: 100%;
+            }
+        }
+
+        @media screen and (max-width: 480px) {
+            #block-info-dialog .text-content {
+                max-height: 30vh;
+            }
         }
     </style>
   </head>
   <body>
-    <div style="text-align: center" class="image-container">
+    <div class="image-container">
       <dialog id="block-info-dialog">
         <button
           class="close-button"
@@ -129,13 +169,11 @@ <h1></h1>
       </dialog>
       <img
         src="$image_data_url"
-        style="max-width: 100%; height: auto"
         alt="Image"
       />
       <svg
         class="blocks-overlay"
-        width="$image_width"
-        height="$image_height"
+        preserveAspectRatio="xMidYMid meet"
       ></svg>
     </div>
     <script src="https://unpkg.com/@popperjs/core@2"></script>
@@ -148,6 +186,20 @@ <h1></h1>
         const blocksById = {};
         const blockInfoDialog = document.querySelector("dialog#block-info-dialog");
 
+        // Handle resizing for responsive SVG
+        function updateSVGSize() {
+          const image = document.querySelector('.image-container img');
+          const svg = document.querySelector('.blocks-overlay');
+
+          // Set SVG dimensions to match the displayed image size
+          svg.setAttribute('width', image.clientWidth);
+          svg.setAttribute('height', image.clientHeight);
+        }
+
+        // Initialize and update on resize
+        window.addEventListener('resize', updateSVGSize);
+        window.addEventListener('load', updateSVGSize);
+
         function blockTypeColor(blockType) {
           return COLORS[BLOCK_TYPES[blockType] % COLORS.length];
         }
@@ -197,11 +249,13 @@ <h1></h1>
         const blocksOverlay = document.querySelector(".blocks-overlay");
         blocksOverlay.innerHTML = traverseAndGenerateSVG(BLOCKS.children[0]);
 
+        // Initialize tippy with responsive settings
         tippy("rect.block", {
             content: (block) => block.getAttribute("data-type"),
             placement: "top-start",
             arrow: false,
             offset: [0, 5],
+            maxWidth: 200,
         });
 
         blocksOverlay.addEventListener("click", (event) => {
@@ -228,7 +282,17 @@ <h1></h1>
             }
             blockInfoDialog.showModal();
         });
-      }; f();
+
+        // Initial sizing
+        updateSVGSize();
+      };
+
+      // Run the function when DOM is loaded
+      if (document.readyState === 'loading') {
+        document.addEventListener('DOMContentLoaded', f);
+      } else {
+        f();
+      }
     </script>
   </body>
-</html>
+</html>
\ No newline at end of file

From f2ddfcdd56572333adf7df3df38550aa1efd8e7c Mon Sep 17 00:00:00 2001
From: Vik Paruchuri <vik.paruchuri@gmail.com>
Date: Tue, 25 Feb 2025 09:59:11 -0500
Subject: [PATCH 26/46] Improve rotated tables

---
 marker/processors/llm/llm_table.py | 45 ++++++++++++++++++++++++++++--
 marker/schema/text/line.py         |  2 +-
 marker/services/__init__.py        |  2 +-
 marker/services/claude.py          |  8 +++---
 4 files changed, 48 insertions(+), 9 deletions(-)

diff --git a/marker/processors/llm/llm_table.py b/marker/processors/llm/llm_table.py
index 6aeb611d..835f8a36 100644
--- a/marker/processors/llm/llm_table.py
+++ b/marker/processors/llm/llm_table.py
@@ -29,16 +29,21 @@ class LLMTableProcessor(BaseLLMComplexBlockProcessor):
         float,
         "The ratio to expand the image by when cropping.",
     ] = 0
+    rotation_max_wh_ratio: Annotated[
+        float,
+        "The maximum width/height ratio for table cells for a table to be considered rotated.",
+    ] = 0.6
     table_rewriting_prompt: Annotated[
         str,
         "The prompt to use for rewriting text.",
         "Default is a string containing the Gemini rewriting prompt."
     ] = """You are a text correction expert specializing in accurately reproducing text from images.
 You will receive an image and an html representation of the table in the image.
-Your task is to correct any errors in the html representation.  The html representation should be as faithful to the original table as possible.
+Your task is to correct any errors in the html representation.  The html representation should be as faithful to the original table as possible.  The table may be rotated, but ensure the html representation is not rotated.  Make sure to include HTML for the full table, including the opening and closing table tags.
 
 Some guidelines:
 - Make sure to reproduce the original values as faithfully as possible.
+- Ensure column headers match the correct column values.
 - If you see any math in a table cell, fence it with the <math> tag.  Block math should be fenced with <math display="block">.
 - Replace any images with a description, like "Image: [description]".
 - Only use the tags th, td, tr, br, span, i, b, math, and table.  Only use the attributes display, style, colspan, and rowspan if necessary.  You can use br to break up text lines in cells.
@@ -47,7 +52,7 @@ class LLMTableProcessor(BaseLLMComplexBlockProcessor):
 **Instructions:**
 1. Carefully examine the provided text block image.
 2. Analyze the html representation of the table.
-3. Write a comparison of the image and the html representation.
+3. Write a comparison of the image and the html representation, paying special attention to the column headers matching the correct column values.
 4. If the html representation is completely correct, or you cannot read the image properly, then write "No corrections needed."  If the html representation has errors, generate the corrected html representation.  Output only either the corrected html representation or "No corrections needed."
 **Example:**
 Input:
@@ -67,7 +72,7 @@ class LLMTableProcessor(BaseLLMComplexBlockProcessor):
 ```
 Output:
 ```html
-Comparison: The image shows a table with 2 rows and 3 columns.  The text and formatting of the html table matches the image.
+Comparison: The image shows a table with 2 rows and 3 columns.  The text and formatting of the html table matches the image.  The column headers match the correct column values.
 No corrections needed.
 ```
 **Input:**
@@ -76,6 +81,35 @@ class LLMTableProcessor(BaseLLMComplexBlockProcessor):
 ```
 """
 
+    def handle_image_rotation(self, children: List[TableCell], image: Image.Image):
+        ratios = [c.polygon.width / c.polygon.height for c in children]
+        if len(ratios) < 2:
+            return image
+
+        is_rotated = all([r < self.rotation_max_wh_ratio for r in ratios])
+        if not is_rotated:
+            return image
+
+        first_col_id = min([c.col_id for c in children])
+        first_col = [c for c in children if c.col_id == first_col_id]
+        first_col_cell = first_col[0]
+
+        last_col_id = max([c.col_id for c in children])
+        if last_col_id == first_col_id:
+            return image
+
+        last_col_cell = [c for c in children if c.col_id == last_col_id][0]
+        cell_diff = first_col_cell.polygon.y_start - last_col_cell.polygon.y_start
+        if cell_diff == 0:
+            return image
+
+        if cell_diff > 0:
+            return image.rotate(270, expand=True)
+        else:
+            return image.rotate(90, expand=True)
+
+
+
     def process_rewriting(self, document: Document, page: PageGroup, block: Table):
         children: List[TableCell] = block.contained_blocks(document, (BlockTypes.TableCell,))
         if not children:
@@ -117,6 +151,7 @@ def process_rewriting(self, document: Document, page: PageGroup, block: Table):
 
             batch_image = block_image.crop(batch_bbox)
             block_html = block.format_cells(document, [], batch_cells)
+            batch_image = self.handle_image_rotation(batch_cells, batch_image)
             batch_parsed_cells = self.rewrite_single_chunk(page, block, block_html, batch_cells, batch_image)
             if batch_parsed_cells is None:
                 return # Error occurred or no corrections needed
@@ -152,6 +187,10 @@ def rewrite_single_chunk(self, page: PageGroup, block: Block, block_html: str, c
             block.update_metadata(llm_error_count=1)
             return
 
+        if not corrected_html.endswith("</table>"):
+            block.update_metadata(llm_error_count=1)
+            return
+
         parsed_cell_text = "".join([cell.text for cell in parsed_cells])
         orig_cell_text = "".join([cell.text for cell in children])
         # Potentially a partial response
diff --git a/marker/schema/text/line.py b/marker/schema/text/line.py
index 4b3bb861..8709b88b 100644
--- a/marker/schema/text/line.py
+++ b/marker/schema/text/line.py
@@ -98,4 +98,4 @@ def merge(self, other: "Line"):
         if self.formats is None:
             self.formats = other.formats
         elif other.formats is not None:
-            self.formats.extend(other.formats)
+            self.formats = list(set(self.formats + other.formats))
diff --git a/marker/services/__init__.py b/marker/services/__init__.py
index 0c7fe1c0..e286ec06 100644
--- a/marker/services/__init__.py
+++ b/marker/services/__init__.py
@@ -11,7 +11,7 @@ class BaseService:
     timeout: Annotated[
         int,
         "The timeout to use for the service."
-    ] = 30
+    ] = 60
     max_retries: Annotated[
         int,
         "The maximum number of retries to use for the service."
diff --git a/marker/services/claude.py b/marker/services/claude.py
index aeef7a2b..cb1055da 100644
--- a/marker/services/claude.py
+++ b/marker/services/claude.py
@@ -7,7 +7,7 @@
 import PIL
 from PIL import Image
 import anthropic
-from anthropic import RateLimitError
+from anthropic import RateLimitError, APITimeoutError
 from pydantic import BaseModel
 
 from marker.schema.blocks import Block
@@ -17,7 +17,7 @@ class ClaudeService(BaseService):
     claude_model_name: Annotated[
         str,
         "The name of the Google model to use for the service."
-    ] = "claude-3-5-sonnet-20241022"
+    ] = "claude-3-7-sonnet-20250219"
     claude_api_key: Annotated[
         str,
         "The Claude API key to use for the service."
@@ -25,7 +25,7 @@ class ClaudeService(BaseService):
     max_claude_tokens: Annotated[
         int,
         "The maximum number of tokens to use for a single Claude request."
-    ] = 4096
+    ] = 8192
 
 
     def img_to_base64(self, img: PIL.Image.Image):
@@ -131,7 +131,7 @@ def __call__(
                 # Extract and validate response
                 response_text = response.content[0].text
                 return self.validate_response(response_text, response_schema)
-            except RateLimitError as e:
+            except (RateLimitError, APITimeoutError) as e:
                 # Rate limit exceeded
                 tries += 1
                 wait_time = tries * 3

From d8a6a1eca10793dcb1bdc0f584089775f31d6b06 Mon Sep 17 00:00:00 2001
From: Vik Paruchuri <vik.paruchuri@gmail.com>
Date: Tue, 25 Feb 2025 10:09:13 -0500
Subject: [PATCH 27/46] Inline math fix

---
 marker/builders/line.py | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/marker/builders/line.py b/marker/builders/line.py
index 752cd28a..55123695 100644
--- a/marker/builders/line.py
+++ b/marker/builders/line.py
@@ -444,6 +444,8 @@ def merge_provider_lines_inline_math(
                     if merge_segment:
                         filtered_merge_lines[line_idx].append(merge_segment)
                     merge_segment = [ml]
+            if merge_segment:
+                filtered_merge_lines[line_idx].append(merge_segment)
 
         # Handle the merging
         already_merged = set()

From 991bb172da5805b297920262659071cf89bfb028 Mon Sep 17 00:00:00 2001
From: Vik Paruchuri <vik.paruchuri@gmail.com>
Date: Tue, 25 Feb 2025 10:36:23 -0500
Subject: [PATCH 28/46] Additional inline math fixes

---
 marker/processors/llm/llm_inlinemath.py | 34 ++++++++++++++++++++++---
 marker/renderers/markdown.py            |  2 +-
 2 files changed, 31 insertions(+), 5 deletions(-)

diff --git a/marker/processors/llm/llm_inlinemath.py b/marker/processors/llm/llm_inlinemath.py
index dd429566..7a1ca127 100644
--- a/marker/processors/llm/llm_inlinemath.py
+++ b/marker/processors/llm/llm_inlinemath.py
@@ -9,7 +9,7 @@
 
 from marker.processors.util import text_to_spans
 from marker.schema import BlockTypes
-from marker.schema.blocks import Block
+from marker.schema.blocks import Block, InlineMath
 from marker.schema.document import Document
 from marker.schema.groups import PageGroup
 from marker.schema.registry import get_block_class
@@ -20,8 +20,14 @@ class LLMInlineMathProcessor(BaseLLMComplexBlockProcessor):
         bool,
         "If True, the inline math will be re-done, otherwise it will be left as is."
     ] = False
+    inlinemath_min_ratio: Annotated[
+        float,
+        "If more than this ratio of blocks are inlinemath blocks, assume everything has math."
+    ] = 0.4
+
+    block_types = (BlockTypes.TextInlineMath,) # Primary block type
+    additional_block_types = (BlockTypes.Text, BlockTypes.Caption, BlockTypes.SectionHeader, BlockTypes.Footnote) # Seconday, can also contain math
 
-    block_types = (BlockTypes.TextInlineMath,)
     text_math_rewriting_prompt = """You are a text correction expert specializing in accurately reproducing text from images.
 You will receive an image of a text block and a set of extracted lines corresponding to the text in the image.
 Your task is to correct any errors in the extracted lines, including math, formatting, and other inaccuracies, and output the corrected lines in a JSON format.
@@ -84,7 +90,7 @@ def rewrite_blocks(self, document: Document):
             return
 
         # Get inline math blocks
-        inline_blocks = [
+        inline_blocks: List[InlineMath] = [
             (page, block)
             for page in document.pages
             for block in page.contained_blocks(document, self.block_types)
@@ -97,7 +103,27 @@ def rewrite_blocks(self, document: Document):
             for block in page.contained_blocks(document, (BlockTypes.Text, BlockTypes.Caption, BlockTypes.SectionHeader, BlockTypes.Footnote))
             if any([b.formats and "math" in b.formats for b in block.contained_blocks(document, (BlockTypes.Line,))])
         ]
-        inference_blocks = inline_blocks + detected_blocks
+
+        # If a page has enough math blocks, assume all blocks can contain math
+        additional_text_blocks = []
+        for page in document.pages:
+            # Check for inline math blocks
+            page_inlinemath_blocks = [im for im in inline_blocks if im[0].page_id == page.page_id]
+            page_detected_blocks = [db for db in detected_blocks if db[0].page_id == page.page_id]
+            math_block_count = len(page_inlinemath_blocks) + len(page_detected_blocks)
+
+            # Find all potential blocks
+            additional_blocks = page.contained_blocks(document, self.additional_block_types + self.block_types)
+
+            # Check if the ratio of math blocks to additional blocks is high enough
+            if math_block_count / len(additional_blocks) < self.inlinemath_min_ratio:
+                continue
+
+            for b in additional_blocks:
+                if b not in detected_blocks and b not in inline_blocks:
+                    additional_text_blocks.append((page, b))
+
+        inference_blocks = inline_blocks + detected_blocks + additional_text_blocks
 
         # Don't show progress if there are no blocks to process
         total_blocks = len(inference_blocks)
diff --git a/marker/renderers/markdown.py b/marker/renderers/markdown.py
index debac078..b524a614 100644
--- a/marker/renderers/markdown.py
+++ b/marker/renderers/markdown.py
@@ -83,7 +83,7 @@ def convert_p(self, el, text, convert_as_inline):
     def convert_math(self, el, text, convert_as_inline):
         inline = el.has_attr('display') and el['display'] == 'inline'
         if inline:
-            return self.inline_math_delimiters[0] + text + self.inline_math_delimiters[1]
+            return " " + self.inline_math_delimiters[0] + text + self.inline_math_delimiters[1] + " "
         else:
             return "\n" + self.block_math_delimiters[0] + text + self.block_math_delimiters[1] + "\n"
 

From 97176d79301a971cf6d7f89020c0767f30e90aec Mon Sep 17 00:00:00 2001
From: Vik Paruchuri <vik.paruchuri@gmail.com>
Date: Tue, 25 Feb 2025 18:59:19 -0500
Subject: [PATCH 29/46] Superscripts

---
 benchmarks/overall/methods/marker.py    | 11 +++-
 benchmarks/overall/methods/olmocr.py    | 88 +++++++++++++++++++++++++
 benchmarks/overall/overall.py           | 23 ++++++-
 benchmarks/overall/registry.py          |  4 +-
 marker/processors/llm/llm_complex.py    |  1 +
 marker/processors/llm/llm_inlinemath.py |  2 +-
 marker/processors/llm/llm_table.py      |  2 +-
 marker/processors/llm/llm_text.py       |  2 +-
 marker/providers/pdf.py                 |  1 +
 marker/renderers/markdown.py            |  4 +-
 marker/schema/text/line.py              |  2 +
 marker/schema/text/span.py              |  4 ++
 12 files changed, 136 insertions(+), 8 deletions(-)
 create mode 100644 benchmarks/overall/methods/olmocr.py

diff --git a/benchmarks/overall/methods/marker.py b/benchmarks/overall/methods/marker.py
index afaafcfc..b17248d5 100644
--- a/benchmarks/overall/methods/marker.py
+++ b/benchmarks/overall/methods/marker.py
@@ -1,7 +1,9 @@
+import os
 import tempfile
 import time
 
 from benchmarks.overall.methods import BaseMethod, BenchmarkResult
+from marker.config.parser import ConfigParser
 from marker.converters.pdf import PdfConverter
 
 
@@ -11,9 +13,16 @@ class MarkerMethod(BaseMethod):
 
     def __call__(self, sample) -> BenchmarkResult:
         pdf_bytes = sample["pdf"]  # This is a single page PDF
+        parser = ConfigParser({
+                "page_range": "0",
+                "disable_tqdm": True,
+                "use_llm": self.use_llm,
+            })
+
         block_converter = PdfConverter(
             artifact_dict=self.model_dict,
-            config={"page_range": [0], "disable_tqdm": True, "use_llm": self.use_llm}
+            config=parser.generate_config_dict(),
+            llm_service=parser.get_llm_service()
         )
 
         with tempfile.NamedTemporaryFile(suffix=".pdf", mode="wb") as f:
diff --git a/benchmarks/overall/methods/olmocr.py b/benchmarks/overall/methods/olmocr.py
new file mode 100644
index 00000000..37f13792
--- /dev/null
+++ b/benchmarks/overall/methods/olmocr.py
@@ -0,0 +1,88 @@
+import base64
+import json
+import tempfile
+import time
+from io import BytesIO
+
+import torch
+from PIL import Image
+
+from benchmarks.overall.methods import BaseMethod, BenchmarkResult
+
+
+def convert_single_page(filename: str, model, processor, device):
+    from olmocr.data.renderpdf import render_pdf_to_base64png
+    from olmocr.prompts import build_finetuning_prompt
+    from olmocr.prompts.anchor import get_anchor_text
+
+    image_base64 = render_pdf_to_base64png(filename, 1, target_longest_image_dim=1024)
+
+    # Build the prompt, using document metadata
+    anchor_text = get_anchor_text(filename, 1, pdf_engine="pdfreport", target_length=4000)
+    prompt = build_finetuning_prompt(anchor_text)
+
+    # Build the full prompt
+    messages = [
+        {
+            "role": "user",
+            "content": [
+                {"type": "text", "text": prompt},
+                {"type": "image_url", "image_url": {"url": f"data:image/png;base64,{image_base64}"}},
+            ],
+        }
+    ]
+
+    # Apply the chat template and processor
+    text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+    main_image = Image.open(BytesIO(base64.b64decode(image_base64)))
+
+    inputs = processor(
+        text=[text],
+        images=[main_image],
+        padding=True,
+        return_tensors="pt",
+    )
+    inputs = {key: value.to(device) for (key, value) in inputs.items()}
+
+    # Generate the output
+    output = model.generate(
+        **inputs,
+        temperature=0.8,
+        max_new_tokens=8192,
+        num_return_sequences=1,
+        do_sample=True,
+    )
+
+    # Decode the output
+    prompt_length = inputs["input_ids"].shape[1]
+    new_tokens = output[:, prompt_length:]
+    text_output = processor.tokenizer.batch_decode(
+        new_tokens, skip_special_tokens=True
+    )[0]
+
+    try:
+        text_output = json.loads(text_output)
+        text = text_output["natural_text"]
+    except Exception:
+        text = text_output.split("natural_text")[1].strip()
+
+    return text
+
+
+class OlmOCRMethod(BaseMethod):
+    olmocr_model: dict = None
+    use_llm: bool = False
+
+    def __call__(self, sample) -> BenchmarkResult:
+        pdf_bytes = sample["pdf"]  # This is a single page PDF
+
+        with tempfile.NamedTemporaryFile(suffix=".pdf", mode="wb") as f:
+            f.write(pdf_bytes)
+            start = time.time()
+            result = convert_single_page(f.name, self.olmocr_model["model"], self.olmocr_model["processor"], self.olmocr_model["model"].device)
+            total = time.time() - start
+
+        return {
+            "markdown": result,
+            "time": total
+        }
diff --git a/benchmarks/overall/overall.py b/benchmarks/overall/overall.py
index 481753e3..68f01368 100644
--- a/benchmarks/overall/overall.py
+++ b/benchmarks/overall/overall.py
@@ -1,11 +1,13 @@
 import json
 import os
+import traceback
 from collections import defaultdict
 from pathlib import Path
 from typing import List
 
 import click
 import datasets
+import torch
 from tqdm import tqdm
 
 from benchmarks.overall.display.dataset import build_dataset
@@ -63,6 +65,7 @@ def get_method_scores(benchmark_dataset: datasets.Dataset, methods: List[str], s
                             averages_by_block_type[method][score_type][gt_block["block_type"]].append(score)
         except Exception as e:
             print(f"Failed to process {idx}: {e}")
+            traceback.print_exc()
             if idx in markdown_by_method:
                 del markdown_by_method[idx]
             continue
@@ -85,6 +88,7 @@ def get_method_scores(benchmark_dataset: datasets.Dataset, methods: List[str], s
 @click.option("--result_path", type=str, default=os.path.join(settings.OUTPUT_DIR, "benchmark", "overall"), help="Output path for results.")
 @click.option("--max_rows", type=int, default=None, help="Maximum number of rows to process.")
 @click.option("--use_llm", is_flag=True, help="Use the LLM model for better marker quality.")
+@click.option("--languages", type=str, help="Comma separated list of languages to use for LLM", default=None)
 def main(
         dataset: str,
         out_dataset: str,
@@ -92,7 +96,8 @@ def main(
         scores: str,
         result_path: str,
         max_rows: int,
-        use_llm: bool
+        use_llm: bool,
+        languages: str
 ):
     out_path = Path(result_path)
     out_path.mkdir(parents=True, exist_ok=True)
@@ -112,7 +117,15 @@ def main(
         if score_type not in SCORE_REGISTRY:
             raise ValueError(f"Score type {score_type} not allowed.  Allowed types are {SCORE_REGISTRY.keys()}")
 
+    if languages:
+        languages = languages.split(",")
+    else:
+        languages = None
+
     benchmark_dataset = datasets.load_dataset(dataset, split="train")
+    if languages:
+        benchmark_dataset = benchmark_dataset.filter(lambda x: x["language"] in languages)
+
     artifacts = {
         "model_dict": create_model_dict(),
         "use_llm": use_llm,
@@ -126,6 +139,14 @@ def main(
     if "llamaparse" in methods:
         artifacts["llamaparse_ds"] = datasets.load_dataset("datalab-to/marker_benchmark_llamaparse", split="train")
 
+    if "olmocr" in methods:
+        from transformers import AutoProcessor, Qwen2VLForConditionalGeneration
+        model = Qwen2VLForConditionalGeneration.from_pretrained("allenai/olmOCR-7B-0225-preview",
+                                                                torch_dtype=torch.bfloat16).eval()
+        processor = AutoProcessor.from_pretrained("Qwen/Qwen2-VL-7B-Instruct")
+        model.to(torch.device("cuda" if torch.cuda.is_available() else "cpu"))
+        artifacts["olmocr_model"] = {"model": model, "processor": processor}
+
     print(f"Running benchmark with methods: {methods} and scores: {score_types}")
     result = get_method_scores(benchmark_dataset, methods, score_types, artifacts, max_rows=max_rows)
 
diff --git a/benchmarks/overall/registry.py b/benchmarks/overall/registry.py
index 02184ad3..eca5d102 100644
--- a/benchmarks/overall/registry.py
+++ b/benchmarks/overall/registry.py
@@ -3,6 +3,7 @@
 from benchmarks.overall.methods.llamaparse import LlamaParseMethod
 from benchmarks.overall.methods.marker import MarkerMethod
 from benchmarks.overall.methods.mathpix import MathpixMethod
+from benchmarks.overall.methods.olmocr import OlmOCRMethod
 from benchmarks.overall.scorers.heuristic import HeuristicScorer
 from benchmarks.overall.scorers.llm import LLMScorer
 
@@ -16,5 +17,6 @@
     "gt": GTMethod,
     "mathpix": MathpixMethod,
     "llamaparse": LlamaParseMethod,
-    "docling": DoclingMethod
+    "docling": DoclingMethod,
+    "olmocr": OlmOCRMethod
 }
\ No newline at end of file
diff --git a/marker/processors/llm/llm_complex.py b/marker/processors/llm/llm_complex.py
index bf856093..59c62415 100644
--- a/marker/processors/llm/llm_complex.py
+++ b/marker/processors/llm/llm_complex.py
@@ -17,6 +17,7 @@ class LLMComplexRegionProcessor(BaseLLMSimpleBlockProcessor):
 
 Formatting should be in markdown, with the following rules:
 - * for italics, ** for bold, and ` for inline code.
+- Use <sup>...</sup> for superscripts.
 - Headers should be formatted with #, with one # for the largest header, and up to 6 for the smallest.
 - Lists should be formatted with either - or 1. for unordered and ordered lists, respectively.
 - Links should be formatted with [text](url).
diff --git a/marker/processors/llm/llm_inlinemath.py b/marker/processors/llm/llm_inlinemath.py
index 7a1ca127..d0d74e2d 100644
--- a/marker/processors/llm/llm_inlinemath.py
+++ b/marker/processors/llm/llm_inlinemath.py
@@ -41,7 +41,7 @@ class LLMInlineMathProcessor(BaseLLMComplexBlockProcessor):
 4. If there are no errors in any of the extracted lines, output "No corrections needed".
 5. For each extracted line, correct any errors, including:
     * Inline math: Ensure all mathematical expressions are correctly formatted and rendered.  Surround them with <math>...</math> tags.
-    * Formatting: Maintain consistent formatting with the text block image, including spacing, indentation, and special characters.
+    * Formatting: Maintain consistent formatting with the text block image, including spacing, indentation, and special characters.  Use the <i>, <b>, <sup>, <sub>, and <span> tags to format the text as needed.
     * Other inaccuracies:  If the image is handwritten then you may correct any spelling errors, or other discrepancies.
 6. Do not remove any formatting i.e bold, italics, math, superscripts, subscripts, etc from the extracted lines unless it is necessary to correct an error.
 7. The number of corrected lines in the output MUST equal the number of extracted lines provided in the input. Do not add or remove lines.
diff --git a/marker/processors/llm/llm_table.py b/marker/processors/llm/llm_table.py
index 835f8a36..264ae29b 100644
--- a/marker/processors/llm/llm_table.py
+++ b/marker/processors/llm/llm_table.py
@@ -46,7 +46,7 @@ class LLMTableProcessor(BaseLLMComplexBlockProcessor):
 - Ensure column headers match the correct column values.
 - If you see any math in a table cell, fence it with the <math> tag.  Block math should be fenced with <math display="block">.
 - Replace any images with a description, like "Image: [description]".
-- Only use the tags th, td, tr, br, span, i, b, math, and table.  Only use the attributes display, style, colspan, and rowspan if necessary.  You can use br to break up text lines in cells.
+- Only use the tags th, td, tr, br, span, sup, sub, i, b, math, and table.  Only use the attributes display, style, colspan, and rowspan if necessary.  You can use br to break up text lines in cells.
 - Make sure the columns and rows match the image faithfully, and are easily readable and interpretable by a human.
 
 **Instructions:**
diff --git a/marker/processors/llm/llm_text.py b/marker/processors/llm/llm_text.py
index ae23619f..c9caa60b 100644
--- a/marker/processors/llm/llm_text.py
+++ b/marker/processors/llm/llm_text.py
@@ -34,7 +34,7 @@ class LLMTextProcessor(BaseLLMSimpleBlockProcessor):
 3. For each extracted line, compare it to the corresponding line in the image.
 4. Correct any errors in the extracted line, including:
     * Inline math: Ensure all mathematical expressions are correctly formatted and rendered.  Use the `<math>` and `</math>` tags to surround inline math properly.  Make sure the opening and closing tags appear in pairs, on the same line.
-    * Formatting: Maintain consistent formatting with the text block image, including spacing, indentation, and special characters.
+    * Formatting: Maintain consistent formatting with the text block image, including spacing, indentation, and special characters.  Use the `<i>`, `<b>`, `<sup>`, `<sub>`, and `<span>` tags to format the text as needed.
     * Other inaccuracies:  If the image is handwritten then you may correct any spelling errors, or other discrepancies.
 5. Do not remove any formatting i.e bold, italics, math, superscripts, subscripts, etc from the extracted lines unless it is necessary to correct an error.  The formatting 
 6. The number of corrected lines in the output MUST equal the number of extracted lines provided in the input. Do not add or remove lines.  There are exactly {line_count} input lines.
diff --git a/marker/providers/pdf.py b/marker/providers/pdf.py
index a63723ff..210b5e0b 100644
--- a/marker/providers/pdf.py
+++ b/marker/providers/pdf.py
@@ -218,6 +218,7 @@ def pdftext_extraction(self, doc: PdfDocument) -> ProviderPageLines:
                                 page_id=page_id,
                                 text_extraction_method="pdftext",
                                 url=span.get("url"),
+                                has_superscript=span.get("superscript", False),
                             )
                         )
                         chars.append(span_chars)
diff --git a/marker/renderers/markdown.py b/marker/renderers/markdown.py
index b524a614..62463273 100644
--- a/marker/renderers/markdown.py
+++ b/marker/renderers/markdown.py
@@ -207,8 +207,8 @@ def md_cls(self):
             heading_style="ATX",
             bullets="-",
             escape_misc=False,
-            escape_underscores=False,
-            escape_asterisks=False,
+            escape_underscores=True,
+            escape_asterisks=True,
             escape_dollars=True,
             sub_symbol="<sub>",
             sup_symbol="<sup>",
diff --git a/marker/schema/text/line.py b/marker/schema/text/line.py
index 8709b88b..5f9134fb 100644
--- a/marker/schema/text/line.py
+++ b/marker/schema/text/line.py
@@ -46,6 +46,8 @@ def formatted_text(self, document):
 
             if block.has_superscript:
                 block_text = re.sub(r"^([0-9\W]+)(.*)", r"<sup>\1</sup>\2", block_text)
+                if "<sup>" not in block_text:
+                    block_text = f"<sup>{block_text}</sup>"
 
             if block.url:
                 block_text = f"<a href='{block.url}'>{block_text}</a>"
diff --git a/marker/schema/text/span.py b/marker/schema/text/span.py
index ab4da1b5..6152fa28 100644
--- a/marker/schema/text/span.py
+++ b/marker/schema/text/span.py
@@ -64,6 +64,10 @@ def assemble_html(self, document, child_blocks, parent_structure):
         if self.has_superscript:
             text = re.sub(r"^([0-9\W]+)(.*)", r"<sup>\1</sup>\2", text)
 
+            # Handle full block superscript
+            if "<sup>" not in text:
+                text = f"<sup>{text}</sup>"
+
         if self.url:
             text = f"<a href='{self.url}'>{text}</a>"
 

From beaad47611b6238a752a72f59fc2b10232a71439 Mon Sep 17 00:00:00 2001
From: Vik Paruchuri <vik.paruchuri@gmail.com>
Date: Tue, 25 Feb 2025 19:56:30 -0500
Subject: [PATCH 30/46] Update inline math

---
 benchmarks/overall/methods/marker.py | 1 +
 1 file changed, 1 insertion(+)

diff --git a/benchmarks/overall/methods/marker.py b/benchmarks/overall/methods/marker.py
index b17248d5..ea8dc8f8 100644
--- a/benchmarks/overall/methods/marker.py
+++ b/benchmarks/overall/methods/marker.py
@@ -17,6 +17,7 @@ def __call__(self, sample) -> BenchmarkResult:
                 "page_range": "0",
                 "disable_tqdm": True,
                 "use_llm": self.use_llm,
+                "redo_inline_math": self.use_llm
             })
 
         block_converter = PdfConverter(

From 790b24495c1ef21506c3af4e82834552227901d4 Mon Sep 17 00:00:00 2001
From: Vik Paruchuri <vik.paruchuri@gmail.com>
Date: Tue, 25 Feb 2025 20:15:58 -0500
Subject: [PATCH 31/46] Fixes to processor

---
 benchmarks/overall/scorers/llm.py       | 2 ++
 marker/processors/llm/llm_inlinemath.py | 2 +-
 2 files changed, 3 insertions(+), 1 deletion(-)

diff --git a/benchmarks/overall/scorers/llm.py b/benchmarks/overall/scorers/llm.py
index 8ee8d138..73b23e12 100644
--- a/benchmarks/overall/scorers/llm.py
+++ b/benchmarks/overall/scorers/llm.py
@@ -48,6 +48,8 @@
 - A 1/5 will have major missing text segments from the markdown or completely unreadable formatting.
 - Use 0/5 if a field isn't applicable, like if the image doesn't contain a table.
 
+If text that is important to the meaning of the document is missing, do not score higher than 3/5.
+
 Output json, like in the example below.
 
 **Example**
diff --git a/marker/processors/llm/llm_inlinemath.py b/marker/processors/llm/llm_inlinemath.py
index d0d74e2d..457e721d 100644
--- a/marker/processors/llm/llm_inlinemath.py
+++ b/marker/processors/llm/llm_inlinemath.py
@@ -116,7 +116,7 @@ def rewrite_blocks(self, document: Document):
             additional_blocks = page.contained_blocks(document, self.additional_block_types + self.block_types)
 
             # Check if the ratio of math blocks to additional blocks is high enough
-            if math_block_count / len(additional_blocks) < self.inlinemath_min_ratio:
+            if math_block_count / max(1, len(additional_blocks)) < self.inlinemath_min_ratio:
                 continue
 
             for b in additional_blocks:

From f8a3c8c2d9aef1bbacece3b5687ae63e2a463a93 Mon Sep 17 00:00:00 2001
From: Vik Paruchuri <vik.paruchuri@gmail.com>
Date: Wed, 26 Feb 2025 07:36:17 -0500
Subject: [PATCH 32/46] Add superscripts

---
 benchmarks/overall/overall.py | 2 +-
 benchmarks/throughput/main.py | 3 ++-
 marker/schema/text/span.py    | 1 +
 pyproject.toml                | 2 +-
 4 files changed, 5 insertions(+), 3 deletions(-)

diff --git a/benchmarks/overall/overall.py b/benchmarks/overall/overall.py
index 68f01368..914296ff 100644
--- a/benchmarks/overall/overall.py
+++ b/benchmarks/overall/overall.py
@@ -27,7 +27,7 @@ def get_method_scores(benchmark_dataset: datasets.Dataset, methods: List[str], s
     averages_by_block_type = defaultdict(lambda: defaultdict(lambda: defaultdict(list)))
     average_times = defaultdict(list)
     markdown_by_method = defaultdict(dict)
-    for idx, sample in tqdm(enumerate(benchmark_dataset), desc="Running benchmark"):
+    for idx, sample in tqdm(enumerate(benchmark_dataset), desc="Running benchmark", total=len(benchmark_dataset)):
         if max_rows is not None and idx >= max_rows:
             break
 
diff --git a/benchmarks/throughput/main.py b/benchmarks/throughput/main.py
index 6e07054b..e6428d35 100644
--- a/benchmarks/throughput/main.py
+++ b/benchmarks/throughput/main.py
@@ -3,6 +3,7 @@
 
 import click
 import pypdfium2 as pdfium
+from tqdm import tqdm
 
 from marker.converters.pdf import PdfConverter
 from marker.models import create_model_dict
@@ -19,7 +20,7 @@ def main(pdf_path):
     torch.cuda.reset_peak_memory_stats()
 
     times = []
-    for i in range(10):
+    for i in tqdm(range(10), desc="Benchmarking"):
         block_converter = PdfConverter(
             artifact_dict=model_dict,
             config={"disable_tqdm": True}
diff --git a/marker/schema/text/span.py b/marker/schema/text/span.py
index 6152fa28..d39a9ae2 100644
--- a/marker/schema/text/span.py
+++ b/marker/schema/text/span.py
@@ -66,6 +66,7 @@ def assemble_html(self, document, child_blocks, parent_structure):
 
             # Handle full block superscript
             if "<sup>" not in text:
+                print(text)
                 text = f"<sup>{text}</sup>"
 
         if self.url:
diff --git a/pyproject.toml b/pyproject.toml
index 32353d92..5b88f2a9 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -28,7 +28,7 @@ ftfy = "^6.1.1"
 rapidfuzz = "^3.8.1"
 surya-ocr = "~0.12.1"
 regex = "^2024.4.28"
-pdftext = "~0.6.0"
+pdftext = "~0.6.1"
 markdownify = "^0.13.1"
 click = "^8.1.7"
 markdown2 = "^2.5.2"

From e2a286143bddf181ba295745002f7d1a2c81473a Mon Sep 17 00:00:00 2001
From: Vik Paruchuri <vik.paruchuri@gmail.com>
Date: Wed, 26 Feb 2025 09:03:40 -0500
Subject: [PATCH 33/46] Fix superscripts

---
 marker/providers/pdf.py    | 8 ++++++--
 marker/schema/text/span.py | 1 -
 2 files changed, 6 insertions(+), 3 deletions(-)

diff --git a/marker/providers/pdf.py b/marker/providers/pdf.py
index 210b5e0b..28c38f5c 100644
--- a/marker/providers/pdf.py
+++ b/marker/providers/pdf.py
@@ -205,10 +205,14 @@ def pdftext_extraction(self, doc: PdfDocument) -> ProviderPageLines:
                         font_size = span["font"]["size"] or 0
                         polygon = PolygonBox.from_bbox(span["bbox"], ensure_nonzero_area=True)
                         span_chars = [Char(char=c['char'], polygon=PolygonBox.from_bbox(c['bbox'], ensure_nonzero_area=True), char_idx=c['char_idx']) for c in span["chars"]]
+                        superscript = span.get("superscript", False)
+                        text = self.normalize_spaces(fix_text(span["text"]))
+                        if superscript:
+                            text = text.strip()
                         spans.append(
                             SpanClass(
                                 polygon=polygon,
-                                text=self.normalize_spaces(fix_text(span["text"])),
+                                text=text,
                                 font=font_name,
                                 font_weight=font_weight,
                                 font_size=font_size,
@@ -218,7 +222,7 @@ def pdftext_extraction(self, doc: PdfDocument) -> ProviderPageLines:
                                 page_id=page_id,
                                 text_extraction_method="pdftext",
                                 url=span.get("url"),
-                                has_superscript=span.get("superscript", False),
+                                has_superscript=superscript,
                             )
                         )
                         chars.append(span_chars)
diff --git a/marker/schema/text/span.py b/marker/schema/text/span.py
index d39a9ae2..6152fa28 100644
--- a/marker/schema/text/span.py
+++ b/marker/schema/text/span.py
@@ -66,7 +66,6 @@ def assemble_html(self, document, child_blocks, parent_structure):
 
             # Handle full block superscript
             if "<sup>" not in text:
-                print(text)
                 text = f"<sup>{text}</sup>"
 
         if self.url:

From ce2ebfacffb3aaee34b52c3399120be38790aa0e Mon Sep 17 00:00:00 2001
From: Vik Paruchuri <vik.paruchuri@gmail.com>
Date: Wed, 26 Feb 2025 09:46:13 -0500
Subject: [PATCH 34/46] Update subscripts

---
 marker/processors/llm/llm_inlinemath.py | 2 +-
 marker/processors/llm/llm_text.py       | 2 +-
 marker/providers/pdf.py                 | 5 ++++-
 marker/schema/text/span.py              | 1 +
 4 files changed, 7 insertions(+), 3 deletions(-)

diff --git a/marker/processors/llm/llm_inlinemath.py b/marker/processors/llm/llm_inlinemath.py
index 457e721d..dace6606 100644
--- a/marker/processors/llm/llm_inlinemath.py
+++ b/marker/processors/llm/llm_inlinemath.py
@@ -41,7 +41,7 @@ class LLMInlineMathProcessor(BaseLLMComplexBlockProcessor):
 4. If there are no errors in any of the extracted lines, output "No corrections needed".
 5. For each extracted line, correct any errors, including:
     * Inline math: Ensure all mathematical expressions are correctly formatted and rendered.  Surround them with <math>...</math> tags.
-    * Formatting: Maintain consistent formatting with the text block image, including spacing, indentation, and special characters.  Use the <i>, <b>, <sup>, <sub>, and <span> tags to format the text as needed.
+    * Formatting: Maintain consistent formatting with the text block image, including spacing, indentation, subscripts/superscripts, and special characters.  Use the <i>, <b>, <sup>, <sub>, and <span> tags to format the text as needed.
     * Other inaccuracies:  If the image is handwritten then you may correct any spelling errors, or other discrepancies.
 6. Do not remove any formatting i.e bold, italics, math, superscripts, subscripts, etc from the extracted lines unless it is necessary to correct an error.
 7. The number of corrected lines in the output MUST equal the number of extracted lines provided in the input. Do not add or remove lines.
diff --git a/marker/processors/llm/llm_text.py b/marker/processors/llm/llm_text.py
index c9caa60b..cc475688 100644
--- a/marker/processors/llm/llm_text.py
+++ b/marker/processors/llm/llm_text.py
@@ -34,7 +34,7 @@ class LLMTextProcessor(BaseLLMSimpleBlockProcessor):
 3. For each extracted line, compare it to the corresponding line in the image.
 4. Correct any errors in the extracted line, including:
     * Inline math: Ensure all mathematical expressions are correctly formatted and rendered.  Use the `<math>` and `</math>` tags to surround inline math properly.  Make sure the opening and closing tags appear in pairs, on the same line.
-    * Formatting: Maintain consistent formatting with the text block image, including spacing, indentation, and special characters.  Use the `<i>`, `<b>`, `<sup>`, `<sub>`, and `<span>` tags to format the text as needed.
+    * Formatting: Maintain consistent formatting with the text block image, including spacing, indentation, subscripts/superscripts, and special characters.  Use the `<i>`, `<b>`, `<sup>`, `<sub>`, and `<span>` tags to format the text as needed.
     * Other inaccuracies:  If the image is handwritten then you may correct any spelling errors, or other discrepancies.
 5. Do not remove any formatting i.e bold, italics, math, superscripts, subscripts, etc from the extracted lines unless it is necessary to correct an error.  The formatting 
 6. The number of corrected lines in the output MUST equal the number of extracted lines provided in the input. Do not add or remove lines.  There are exactly {line_count} input lines.
diff --git a/marker/providers/pdf.py b/marker/providers/pdf.py
index 28c38f5c..7af0b44c 100644
--- a/marker/providers/pdf.py
+++ b/marker/providers/pdf.py
@@ -206,9 +206,11 @@ def pdftext_extraction(self, doc: PdfDocument) -> ProviderPageLines:
                         polygon = PolygonBox.from_bbox(span["bbox"], ensure_nonzero_area=True)
                         span_chars = [Char(char=c['char'], polygon=PolygonBox.from_bbox(c['bbox'], ensure_nonzero_area=True), char_idx=c['char_idx']) for c in span["chars"]]
                         superscript = span.get("superscript", False)
+                        subscript = span.get("subscript", False)
                         text = self.normalize_spaces(fix_text(span["text"]))
-                        if superscript:
+                        if superscript or superscript:
                             text = text.strip()
+
                         spans.append(
                             SpanClass(
                                 polygon=polygon,
@@ -223,6 +225,7 @@ def pdftext_extraction(self, doc: PdfDocument) -> ProviderPageLines:
                                 text_extraction_method="pdftext",
                                 url=span.get("url"),
                                 has_superscript=superscript,
+                                has_subscript=subscript
                             )
                         )
                         chars.append(span_chars)
diff --git a/marker/schema/text/span.py b/marker/schema/text/span.py
index 6152fa28..a8f7e0d6 100644
--- a/marker/schema/text/span.py
+++ b/marker/schema/text/span.py
@@ -24,6 +24,7 @@ class Span(Block):
     maximum_position: int
     formats: List[Literal['plain', 'math', 'chemical', 'bold', 'italic']]
     has_superscript: bool = False
+    has_subscript: bool = False
     url: Optional[str] = None
 
     @property

From 92c395f82a1a85bd04ff1a4d19a7275f771d5eef Mon Sep 17 00:00:00 2001
From: Vik Paruchuri <vik.paruchuri@gmail.com>
Date: Wed, 26 Feb 2025 09:59:49 -0500
Subject: [PATCH 35/46] Fix superscripts with llm mode

---
 marker/processors/util.py | 13 +++++++++++--
 1 file changed, 11 insertions(+), 2 deletions(-)

diff --git a/marker/processors/util.py b/marker/processors/util.py
index e26234df..d3bd2cf9 100644
--- a/marker/processors/util.py
+++ b/marker/processors/util.py
@@ -37,6 +37,8 @@ def add_math_spans_to_line(corrected_text: str, text_line: Line, page: PageGroup
                 url=span.get('url'),
                 page_id=text_line.page_id,
                 text_extraction_method="gemini",
+                has_superscript=span["has_superscript"],
+                has_subscript=span["has_subscript"]
             )
         )
         text_line.structure.append(span_block.id)
@@ -49,6 +51,9 @@ def text_to_spans(text):
         'b': 'bold',
         'i': 'italic',
         'math': 'math',
+        'sub': 'plain',
+        'sup': 'plain',
+        'span': 'plain'
     }
     spans = []
 
@@ -65,13 +70,17 @@ def text_to_spans(text):
             spans.append({
                 'type': tag_types[element.name],
                 'content': text,
-                'url': url
+                'url': url,
+                "has_superscript": element.name == "sup",
+                "has_subscript": element.name == "sub"
             })
         elif element.string:
             spans.append({
                 'type': 'plain',
                 'content': element.string,
-                'url': url
+                'url': url,
+                "has_superscript": False,
+                "has_subscript": False
             })
 
     return spans
\ No newline at end of file

From c5d433a2e30e914c83f581fbb0502b97f0dade11 Mon Sep 17 00:00:00 2001
From: Vik Paruchuri <vik.paruchuri@gmail.com>
Date: Wed, 26 Feb 2025 10:02:33 -0500
Subject: [PATCH 36/46] Bump pdftext

---
 poetry.lock | 52 +++++++++++++++++++++++++++-------------------------
 1 file changed, 27 insertions(+), 25 deletions(-)

diff --git a/poetry.lock b/poetry.lock
index f1e90012..e5f79050 100644
--- a/poetry.lock
+++ b/poetry.lock
@@ -1433,16 +1433,18 @@ requests = ["requests (>=2.20.0,<3.0.0.dev0)"]
 
 [[package]]
 name = "google-genai"
-version = "1.2.0"
+version = "1.3.0"
 description = "GenAI Python SDK"
 optional = false
 python-versions = ">=3.9"
 files = [
-    {file = "google_genai-1.2.0-py3-none-any.whl", hash = "sha256:609d61bee73f1a6ae5b47e9c7dd4b469d50318f050c5ceacf835b0f80f79d2d9"},
+    {file = "google_genai-1.3.0-py3-none-any.whl", hash = "sha256:daa8934addb701ff5863d80f5eed278b33cac5dcb41e8eba9363fce8827c308b"},
+    {file = "google_genai-1.3.0.tar.gz", hash = "sha256:7b365a767474becc899bb2f1a38bcdc4967f20645cbf9b7781bd38ad1a59c25b"},
 ]
 
 [package.dependencies]
 google-auth = ">=2.14.1,<3.0.0dev"
+httpx = ">=0.28.1,<1.0.0dev"
 pydantic = ">=2.0.0,<3.0.0dev"
 requests = ">=2.28.1,<3.0.0dev"
 typing-extensions = ">=4.11.0,<5.0.0dev"
@@ -3287,13 +3289,13 @@ testing = ["docopt", "pytest"]
 
 [[package]]
 name = "pdftext"
-version = "0.6.0"
+version = "0.6.1"
 description = "Extract structured text from pdfs quickly"
 optional = false
 python-versions = "<4.0,>=3.10"
 files = [
-    {file = "pdftext-0.6.0-py3-none-any.whl", hash = "sha256:13f3559a6a24d6a5dbacc8b31e094cbe1fc0b284e6fe81df68bb290ef719f245"},
-    {file = "pdftext-0.6.0.tar.gz", hash = "sha256:0de0a4a5a448cdf28aea30706b3b79013aa8e679488e4d7a57cf69407ab9fb46"},
+    {file = "pdftext-0.6.1-py3-none-any.whl", hash = "sha256:9c437a05262277dede2f6953eebc7b46d7393bb11ee373267814af2aa5e02e4d"},
+    {file = "pdftext-0.6.1.tar.gz", hash = "sha256:ffec41064804e157b48b76c834051ed7a5aa456257b78d9b87a5e8f54cebe307"},
 ]
 
 [package.dependencies]
@@ -4769,26 +4771,26 @@ pyasn1 = ">=0.1.3"
 
 [[package]]
 name = "safetensors"
-version = "0.5.2"
+version = "0.5.3"
 description = ""
 optional = false
 python-versions = ">=3.7"
 files = [
-    {file = "safetensors-0.5.2-cp38-abi3-macosx_10_12_x86_64.whl", hash = "sha256:45b6092997ceb8aa3801693781a71a99909ab9cc776fbc3fa9322d29b1d3bef2"},
-    {file = "safetensors-0.5.2-cp38-abi3-macosx_11_0_arm64.whl", hash = "sha256:6d0d6a8ee2215a440e1296b843edf44fd377b055ba350eaba74655a2fe2c4bae"},
-    {file = "safetensors-0.5.2-cp38-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:86016d40bcaa3bcc9a56cd74d97e654b5f4f4abe42b038c71e4f00a089c4526c"},
-    {file = "safetensors-0.5.2-cp38-abi3-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:990833f70a5f9c7d3fc82c94507f03179930ff7d00941c287f73b6fcbf67f19e"},
-    {file = "safetensors-0.5.2-cp38-abi3-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:3dfa7c2f3fe55db34eba90c29df94bcdac4821043fc391cb5d082d9922013869"},
-    {file = "safetensors-0.5.2-cp38-abi3-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:46ff2116150ae70a4e9c490d2ab6b6e1b1b93f25e520e540abe1b81b48560c3a"},
-    {file = "safetensors-0.5.2-cp38-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:3ab696dfdc060caffb61dbe4066b86419107a24c804a4e373ba59be699ebd8d5"},
-    {file = "safetensors-0.5.2-cp38-abi3-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:03c937100f38c9ff4c1507abea9928a6a9b02c9c1c9c3609ed4fb2bf413d4975"},
-    {file = "safetensors-0.5.2-cp38-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:a00e737948791b94dad83cf0eafc09a02c4d8c2171a239e8c8572fe04e25960e"},
-    {file = "safetensors-0.5.2-cp38-abi3-musllinux_1_2_armv7l.whl", hash = "sha256:d3a06fae62418ec8e5c635b61a8086032c9e281f16c63c3af46a6efbab33156f"},
-    {file = "safetensors-0.5.2-cp38-abi3-musllinux_1_2_i686.whl", hash = "sha256:1506e4c2eda1431099cebe9abf6c76853e95d0b7a95addceaa74c6019c65d8cf"},
-    {file = "safetensors-0.5.2-cp38-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:5c5b5d9da594f638a259fca766046f44c97244cc7ab8bef161b3e80d04becc76"},
-    {file = "safetensors-0.5.2-cp38-abi3-win32.whl", hash = "sha256:fe55c039d97090d1f85277d402954dd6ad27f63034fa81985a9cc59655ac3ee2"},
-    {file = "safetensors-0.5.2-cp38-abi3-win_amd64.whl", hash = "sha256:78abdddd03a406646107f973c7843276e7b64e5e32623529dc17f3d94a20f589"},
-    {file = "safetensors-0.5.2.tar.gz", hash = "sha256:cb4a8d98ba12fa016f4241932b1fc5e702e5143f5374bba0bbcf7ddc1c4cf2b8"},
+    {file = "safetensors-0.5.3-cp38-abi3-macosx_10_12_x86_64.whl", hash = "sha256:bd20eb133db8ed15b40110b7c00c6df51655a2998132193de2f75f72d99c7073"},
+    {file = "safetensors-0.5.3-cp38-abi3-macosx_11_0_arm64.whl", hash = "sha256:21d01c14ff6c415c485616b8b0bf961c46b3b343ca59110d38d744e577f9cce7"},
+    {file = "safetensors-0.5.3-cp38-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:11bce6164887cd491ca75c2326a113ba934be596e22b28b1742ce27b1d076467"},
+    {file = "safetensors-0.5.3-cp38-abi3-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:4a243be3590bc3301c821da7a18d87224ef35cbd3e5f5727e4e0728b8172411e"},
+    {file = "safetensors-0.5.3-cp38-abi3-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:8bd84b12b1670a6f8e50f01e28156422a2bc07fb16fc4e98bded13039d688a0d"},
+    {file = "safetensors-0.5.3-cp38-abi3-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:391ac8cab7c829452175f871fcaf414aa1e292b5448bd02620f675a7f3e7abb9"},
+    {file = "safetensors-0.5.3-cp38-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:cead1fa41fc54b1e61089fa57452e8834f798cb1dc7a09ba3524f1eb08e0317a"},
+    {file = "safetensors-0.5.3-cp38-abi3-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:1077f3e94182d72618357b04b5ced540ceb71c8a813d3319f1aba448e68a770d"},
+    {file = "safetensors-0.5.3-cp38-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:799021e78287bac619c7b3f3606730a22da4cda27759ddf55d37c8db7511c74b"},
+    {file = "safetensors-0.5.3-cp38-abi3-musllinux_1_2_armv7l.whl", hash = "sha256:df26da01aaac504334644e1b7642fa000bfec820e7cef83aeac4e355e03195ff"},
+    {file = "safetensors-0.5.3-cp38-abi3-musllinux_1_2_i686.whl", hash = "sha256:32c3ef2d7af8b9f52ff685ed0bc43913cdcde135089ae322ee576de93eae5135"},
+    {file = "safetensors-0.5.3-cp38-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:37f1521be045e56fc2b54c606d4455573e717b2d887c579ee1dbba5f868ece04"},
+    {file = "safetensors-0.5.3-cp38-abi3-win32.whl", hash = "sha256:cfc0ec0846dcf6763b0ed3d1846ff36008c6e7290683b61616c4b040f6a54ace"},
+    {file = "safetensors-0.5.3-cp38-abi3-win_amd64.whl", hash = "sha256:836cbbc320b47e80acd40e44c8682db0e8ad7123209f69b093def21ec7cafd11"},
+    {file = "safetensors-0.5.3.tar.gz", hash = "sha256:b6b0d6ecacec39a4fdd99cc19f4576f5219ce858e6fd8dbe7609df0b8dc56965"},
 ]
 
 [package.extras]
@@ -4939,13 +4941,13 @@ win32 = ["pywin32"]
 
 [[package]]
 name = "setuptools"
-version = "75.8.0"
+version = "75.8.1"
 description = "Easily download, build, install, upgrade, and uninstall Python packages"
 optional = false
 python-versions = ">=3.9"
 files = [
-    {file = "setuptools-75.8.0-py3-none-any.whl", hash = "sha256:e3982f444617239225d675215d51f6ba05f845d4eec313da4418fdbb56fb27e3"},
-    {file = "setuptools-75.8.0.tar.gz", hash = "sha256:c5afc8f407c626b8313a86e10311dd3f661c6cd9c09d4bf8c15c0e11f9f2b0e6"},
+    {file = "setuptools-75.8.1-py3-none-any.whl", hash = "sha256:3bc32c0b84c643299ca94e77f834730f126efd621de0cc1de64119e0e17dab1f"},
+    {file = "setuptools-75.8.1.tar.gz", hash = "sha256:65fb779a8f28895242923582eadca2337285f0891c2c9e160754df917c3d2530"},
 ]
 
 [package.extras]
@@ -6107,4 +6109,4 @@ test = ["pytest"]
 [metadata]
 lock-version = "2.0"
 python-versions = "^3.10"
-content-hash = "0a6d5e377e87278aa87ec01a22dfcb38fa9e9083136fe73d3a9d1bec02683dae"
+content-hash = "12c3ac5c7dfd66a41559c675147d24de82ab83feceb1a0fc150772d1fc8864c0"

From b10040883a949cbf90878a5e9735d6abf796a667 Mon Sep 17 00:00:00 2001
From: Vik Paruchuri <vik.paruchuri@gmail.com>
Date: Wed, 26 Feb 2025 10:42:54 -0500
Subject: [PATCH 37/46] Update llm services

---
 benchmarks/overall/methods/marker.py | 4 +++-
 benchmarks/overall/scorers/llm.py    | 7 +++++--
 marker/services/__init__.py          | 2 +-
 marker/services/gemini.py            | 2 +-
 4 files changed, 10 insertions(+), 5 deletions(-)

diff --git a/benchmarks/overall/methods/marker.py b/benchmarks/overall/methods/marker.py
index ea8dc8f8..97239e95 100644
--- a/benchmarks/overall/methods/marker.py
+++ b/benchmarks/overall/methods/marker.py
@@ -17,7 +17,9 @@ def __call__(self, sample) -> BenchmarkResult:
                 "page_range": "0",
                 "disable_tqdm": True,
                 "use_llm": self.use_llm,
-                "redo_inline_math": self.use_llm
+                "redo_inline_math": self.use_llm,
+                "llm_service": "marker.services.vertex.GoogleVertexService",
+                "vertex_project_id": os.getenv("VERTEX_PROJECT_ID"),
             })
 
         block_converter = PdfConverter(
diff --git a/benchmarks/overall/scorers/llm.py b/benchmarks/overall/scorers/llm.py
index 73b23e12..40cd0e82 100644
--- a/benchmarks/overall/scorers/llm.py
+++ b/benchmarks/overall/scorers/llm.py
@@ -1,4 +1,5 @@
 import json
+import os
 import tempfile
 import time
 from typing import List
@@ -126,8 +127,10 @@ def llm_rater(self, img: Image.Image, markdown: str) -> BlockScores:
 
     def llm_response_wrapper(self, prompt, response_schema, depth=0):
         client = genai.Client(
-            api_key=settings.GOOGLE_API_KEY,
-            http_options={"timeout": 60000}
+            http_options={"timeout": 60000},
+            vertexai=True,
+            project=os.getenv("VERTEX_PROJECT_ID"),
+            location=os.getenv("VERTEX_LOCATION"),
         )
         try:
             responses = client.models.generate_content(
diff --git a/marker/services/__init__.py b/marker/services/__init__.py
index e286ec06..0c7fe1c0 100644
--- a/marker/services/__init__.py
+++ b/marker/services/__init__.py
@@ -11,7 +11,7 @@ class BaseService:
     timeout: Annotated[
         int,
         "The timeout to use for the service."
-    ] = 60
+    ] = 30
     max_retries: Annotated[
         int,
         "The maximum number of retries to use for the service."
diff --git a/marker/services/gemini.py b/marker/services/gemini.py
index d393a8b8..990a6b4a 100644
--- a/marker/services/gemini.py
+++ b/marker/services/gemini.py
@@ -64,7 +64,7 @@ def __call__(
                 block.update_metadata(llm_tokens_used=total_tokens, llm_request_count=1)
                 return json.loads(output)
             except APIError as e:
-                if e.code == 429:
+                if e.code in [429, 443, 503]:
                     # Rate limit exceeded
                     tries += 1
                     wait_time = tries * 3

From 4962303a19219adb571725303328a6deef6808db Mon Sep 17 00:00:00 2001
From: Vik Paruchuri <vik.paruchuri@gmail.com>
Date: Thu, 27 Feb 2025 19:33:57 -0500
Subject: [PATCH 38/46] Minor benchmark updates

---
 benchmarks/overall/display/dataset.py   |  7 ++-
 benchmarks/overall/elo.py               | 75 ++++++++++++-------------
 benchmarks/overall/methods/olmocr.py    |  5 +-
 benchmarks/overall/overall.py           |  8 ++-
 benchmarks/overall/scorers/heuristic.py |  9 +++
 benchmarks/overall/scorers/llm.py       | 10 +++-
 marker/config/parser.py                 |  2 +-
 marker/services/__init__.py             |  2 +-
 8 files changed, 72 insertions(+), 46 deletions(-)

diff --git a/benchmarks/overall/display/dataset.py b/benchmarks/overall/display/dataset.py
index e9fcabdd..b38bcf9f 100644
--- a/benchmarks/overall/display/dataset.py
+++ b/benchmarks/overall/display/dataset.py
@@ -29,7 +29,12 @@ def build_dataset(bench_dataset: datasets.Dataset, result: FullResult, score_typ
 
             method_cls = METHOD_REGISTRY[method]()
             md = result["markdown"][idx][method]
-            method_img = method_cls.render(result["markdown"][idx][method])
+            try:
+                method_img = method_cls.render(result["markdown"][idx][method])
+            except Exception as e:
+                # This can happen when the markdown is None
+                method_img = PIL.Image.new("RGB", (200, 200))
+
             row[f"{method}_md"] = md
             row[f"{method}_img"] = method_img
 
diff --git a/benchmarks/overall/elo.py b/benchmarks/overall/elo.py
index 9eea3b55..72b298ce 100644
--- a/benchmarks/overall/elo.py
+++ b/benchmarks/overall/elo.py
@@ -1,9 +1,12 @@
 import json
 import random
 import time
+import os
 from dataclasses import dataclass
 from typing import List, Dict, Tuple, Literal
 from PIL import Image
+from collections import defaultdict
+import tabulate
 
 import click
 import datasets
@@ -48,7 +51,7 @@
 
 Notes on scoring:
 - Perfect markdown will include all of the important text from the image, and the formatting will be correct (minor mistakes okay).  It's okay to omit some text that isn't important to the meaning, like page numbers and chapter headings.  If the entire page is an image, it's okay if the markdown is just a link to the image, unless the image would be better represented as text.
-- Bad markdown will have major missing text segments from the markdown or completely unreadable formatting.
+- Bad markdown will have major missing text segments from the markdown or completely unreadable formatting.  It may also have key values that are different from the values in the image.
 
 Output json, like in the example below.
 
@@ -63,15 +66,15 @@
 ```markdown
 # Section 1
 This is some markdown extracted from a document.  Here is a block equation:
-$$\frac{ab \cdot x^5 + x^2 + 2 \cdot x + 123}{t}$$
+$$\frac{ab \cdot x^5 + x^2 + 2 \cdot x + 124}{t}$$
 ```
 Output
 ```json
 {
     "image_description": "In the image, there is a section header 'Section 1', followed by some text and a block equation.",
     "version_a_description": "In the markdown, there is a section header 'Section 1', followed by some text and a block equation.",
-    "version_b_description": "In the markdown, there is a section header 'Section 1', followed by some text and a block equation.  The formatting in version b is slightly different from the image.",
-    "comparison": "Version A is better than version B.  The text and formatting in version A matches the image better than version B.",
+    "version_b_description": "In the markdown, there is a section header 'Section 1', followed by some text and a block equation.  The formatting in version b is slightly different from the image.  The value 124 is also different from the image.",
+    "comparison": "Version A is better than version B.  The text and formatting in version A matches the image better than version B.  Version B also has an incorrect value.",
     "winner": "version_a",
 }
 ```
@@ -105,6 +108,11 @@ def __call__(
         version_a: str,
         version_b: str
     ) -> str | None:
+        if version_a is None and version_b is not None:
+            return "version_b"
+        elif version_b is None and version_a is not None:
+            return "version_a"
+
         hydrated_prompt = rating_prompt.replace("{{version_a}}", version_a).replace("{{version_b}}", version_b)
         try:
             rating = self.llm_rater(img, hydrated_prompt)
@@ -128,12 +136,14 @@ def llm_response_wrapper(
         response_schema,
     ):
         client = genai.Client(
-            api_key=settings.GOOGLE_API_KEY,
-            http_options={"timeout": 60000}
+            http_options={"timeout": 60000},
+            vertexai=True,
+            project=os.getenv("VERTEX_PROJECT_ID"),
+            location=os.getenv("VERTEX_LOCATION"),
         )
         try:
             responses = client.models.generate_content(
-                model="gemini-2.0-flash",
+                model="gemini-2.0-flash-001",
                 contents=prompt,
                 config={
                     "temperature": 0,
@@ -150,35 +160,19 @@ def llm_response_wrapper(
             print(f"Error: {e}")
             return
 
-@dataclass
-class Method:
-    name: str
-    rating: float = 1500
-    k_factor: float = 32
-
-
-class EloSystem:
-    def __init__(self, player_names: List[str]):
-        self.methods = {name: Method(name) for name in player_names}
-
-    def expected_score(self, rating_a: float, rating_b: float) -> float:
-        return 1 / (1 + 10 ** ((rating_b - rating_a) / 400))
-
-    def update_ratings(self, winner: str, loser: str) -> Tuple[float, float]:
-        method_a = self.methods[winner]
-        method_b = self.methods[loser]
-
-        expected_a = self.expected_score(method_a.rating, method_b.rating)
-        expected_b = self.expected_score(method_b.rating, method_a.rating)
-
-        # Winner gets score of 1, loser gets 0
-        method_a.rating += method_a.k_factor * (1 - expected_a)
-        method_b.rating += method_b.k_factor * (0 - expected_b)
 
-        return method_a.rating, method_b.rating
+def display_win_rates_table(win_rates: dict):
+    table = []
+    headers = ["Method A", "Method B", "Wins", "Losses", "Win %"]
+    for method_a, method_b_dict in win_rates.items():
+        row = [method_a]
+        for method_b, results in method_b_dict.items():
+            row = [method_a, method_b, results["win"], results["loss"], (results["win"] / (results["win"] + results["loss"])) * 100]
+            table.append(row)
+    print(tabulate.tabulate(table, headers=headers, tablefmt="pretty"))
 
 
-@click.command("Calculate ELO scores for document conversion methods")
+@click.command("Calculate win rates for document conversion methods")
 @click.argument("dataset", type=str)
 @click.option("--methods", type=str, help="List of methods to compare: comma separated like marker,mathpix")
 @click.option("--row_samples", type=int, default=2, help="Number of samples per row")
@@ -191,10 +185,10 @@ def main(
 ):
     ds = datasets.load_dataset(dataset, split="train")
     method_lst = methods.split(",")
-    elo = EloSystem(method_lst)
+    win_rates = {m: defaultdict(lambda: defaultdict(int)) for m in method_lst}
     comparer = Comparer()
 
-    for i in tqdm(range(min(len(ds), max_rows)), desc="Calculating ELO"):
+    for i in tqdm(range(min(len(ds), max_rows)), desc="Calculating win rates..."):
         row = ds[i]
         # Avoid any bias in ordering
         random.shuffle(method_lst)
@@ -211,14 +205,15 @@ def main(
                     continue
 
                 if winner == "version_a":
-                    elo.update_ratings(method_a, method_b)
+                    win_rates[method_a][method_b]["win"] += 1
+                    win_rates[method_b][method_a]["loss"] += 1
                 else:
-                    elo.update_ratings(method_b, method_a)
+                    win_rates[method_b][method_a]["win"] += 1
+                    win_rates[method_a][method_b]["loss"] += 1
         if i % 10 == 0:
-            print(elo.methods)
+            display_win_rates_table(win_rates)
 
-    # Print out ratings
-    print(elo.methods)
+    display_win_rates_table(win_rates)
 
 
 if __name__ == "__main__":
diff --git a/benchmarks/overall/methods/olmocr.py b/benchmarks/overall/methods/olmocr.py
index 37f13792..28198899 100644
--- a/benchmarks/overall/methods/olmocr.py
+++ b/benchmarks/overall/methods/olmocr.py
@@ -64,7 +64,10 @@ def convert_single_page(filename: str, model, processor, device):
         text_output = json.loads(text_output)
         text = text_output["natural_text"]
     except Exception:
-        text = text_output.split("natural_text")[1].strip()
+        try:
+            text = text_output.split("natural_text")[1].strip()
+        except Exception:
+            text = ""
 
     return text
 
diff --git a/benchmarks/overall/overall.py b/benchmarks/overall/overall.py
index 914296ff..9c1ccdbe 100644
--- a/benchmarks/overall/overall.py
+++ b/benchmarks/overall/overall.py
@@ -27,7 +27,10 @@ def get_method_scores(benchmark_dataset: datasets.Dataset, methods: List[str], s
     averages_by_block_type = defaultdict(lambda: defaultdict(lambda: defaultdict(list)))
     average_times = defaultdict(list)
     markdown_by_method = defaultdict(dict)
-    for idx, sample in tqdm(enumerate(benchmark_dataset), desc="Running benchmark", total=len(benchmark_dataset)):
+    total_rows = len(benchmark_dataset)
+    if max_rows:
+        total_rows = min(max_rows, total_rows)
+    for idx, sample in tqdm(enumerate(benchmark_dataset), desc="Running benchmark", total=total_rows):
         if max_rows is not None and idx >= max_rows:
             break
 
@@ -44,6 +47,9 @@ def get_method_scores(benchmark_dataset: datasets.Dataset, methods: List[str], s
                 method_cls = METHOD_REGISTRY[method](**artifacts)
                 method_info = method_cls(sample)
                 method_md = method_info["markdown"]
+                if method_md is None:
+                    method_md = "" # Avoid None values
+
                 average_times[method].append(method_info["time"])
                 markdown_by_method[idx][method] = method_md
 
diff --git a/benchmarks/overall/scorers/heuristic.py b/benchmarks/overall/scorers/heuristic.py
index ac1bf0e0..b457d94b 100644
--- a/benchmarks/overall/scorers/heuristic.py
+++ b/benchmarks/overall/scorers/heuristic.py
@@ -9,6 +9,15 @@
 
 class HeuristicScorer(BaseScorer):
     def __call__(self, sample, gt_markdown: List[str], method_markdown: str) -> BlockScores:
+        if not method_markdown:
+            return {
+                "score": 0,
+                "specific_scores": {
+                    "order": 0,
+                    "by_block": [0] * len(gt_markdown)
+                }
+            }
+
         # Standardize inputs
         gt_markdown = [self.clean_input(block) for block in gt_markdown]
         method_markdown = self.clean_input(method_markdown)
diff --git a/benchmarks/overall/scorers/llm.py b/benchmarks/overall/scorers/llm.py
index 40cd0e82..00ff4031 100644
--- a/benchmarks/overall/scorers/llm.py
+++ b/benchmarks/overall/scorers/llm.py
@@ -106,6 +106,14 @@ def __call__(self, sample, gt_markdown: List[str], markdown: str) -> BlockScores
 
 
     def llm_rater(self, img: Image.Image, markdown: str) -> BlockScores:
+        if not markdown:
+            null_scores = {k: 1 for k in score_keys}
+            text_scores = {k: "" for k in text_keys}
+            null_scores.update(text_scores)
+            return {
+                "score": 1,
+                "specific_scores": null_scores
+            }
         req_keys = text_keys + score_keys
         properties = {}
         for key in req_keys:
@@ -134,7 +142,7 @@ def llm_response_wrapper(self, prompt, response_schema, depth=0):
         )
         try:
             responses = client.models.generate_content(
-                model="gemini-2.0-flash",
+                model="gemini-2.0-flash-001",
                 contents=prompt,
                 config={
                     "temperature": 0,
diff --git a/marker/config/parser.py b/marker/config/parser.py
index 1bff73e3..676de7f2 100644
--- a/marker/config/parser.py
+++ b/marker/config/parser.py
@@ -39,7 +39,7 @@ def common_options(fn):
         fn = click.option("--languages", type=str, default=None, help="Comma separated list of languages to use for OCR.")(fn)
 
         # we put common options here
-        fn = click.option("--use_llm", is_flag=True, default=False, help="Enable higher quality processing with LLMs.")(fn)
+        fn = click.option("--use_llm", default=False, help="Enable higher quality processing with LLMs.")(fn)
         fn = click.option("--converter_cls", type=str, default=None, help="Converter class to use.  Defaults to PDF converter.")(fn)
         fn = click.option("--llm_service", type=str, default=None, help="LLM service to use - should be full import path, like marker.services.gemini.GoogleGeminiService")(fn)
 
diff --git a/marker/services/__init__.py b/marker/services/__init__.py
index 0c7fe1c0..1b89d713 100644
--- a/marker/services/__init__.py
+++ b/marker/services/__init__.py
@@ -15,7 +15,7 @@ class BaseService:
     max_retries: Annotated[
         int,
         "The maximum number of retries to use for the service."
-    ] = 3
+    ] = 2
 
     def __init__(self, config: Optional[BaseModel | dict] = None):
         assign_config(self, config)

From 4c29e753f9f006c36473e7aa7343960e1eb25fd8 Mon Sep 17 00:00:00 2001
From: Vik Paruchuri <vik.paruchuri@gmail.com>
Date: Fri, 28 Feb 2025 10:40:46 -0500
Subject: [PATCH 39/46] Enable config in CLI app

---
 marker/processors/llm/llm_equation.py   |   7 +-
 marker/processors/llm/llm_inlinemath.py | 102 +++++++-----------------
 marker/processors/llm/llm_text.py       |   2 +-
 marker/schema/blocks/base.py            |   5 ++
 marker/schema/blocks/caption.py         |   7 ++
 marker/schema/blocks/footnote.py        |   7 ++
 marker/schema/blocks/inlinemath.py      |   4 +
 marker/schema/blocks/listitem.py        |   5 ++
 marker/schema/blocks/sectionheader.py   |   4 +
 marker/schema/blocks/text.py            |   4 +-
 marker/scripts/run_streamlit_app.py     |   4 +
 marker/scripts/streamlit_app.py         |  31 ++++++-
 12 files changed, 104 insertions(+), 78 deletions(-)

diff --git a/marker/processors/llm/llm_equation.py b/marker/processors/llm/llm_equation.py
index e1568842..06b616a1 100644
--- a/marker/processors/llm/llm_equation.py
+++ b/marker/processors/llm/llm_equation.py
@@ -17,6 +17,10 @@ class LLMEquationProcessor(BaseLLMSimpleBlockProcessor):
         float,
         "The ratio to expand the image by when cropping.",
     ] = 0.05 # Equations sometimes get bboxes that are too tight
+    redo_inline_math: Annotated[
+        bool,
+        "Whether to redo inline math blocks.",
+    ] = False
     equation_latex_prompt: Annotated[
         str,
         "The prompt to use for generating LaTeX from equations.",
@@ -65,7 +69,8 @@ def inference_blocks(self, document: Document) -> List[BlockData]:
         for block_data in blocks:
             block = block_data["block"]
             page = block_data["page"]
-            if block.polygon.height / page.polygon.height < self.min_equation_height:
+            # If we redo inline math, we redo all equations
+            if block.polygon.height / page.polygon.height < self.min_equation_height and not self.redo_inline_math:
                 continue
             out_blocks.append(block_data)
         return out_blocks
diff --git a/marker/processors/llm/llm_inlinemath.py b/marker/processors/llm/llm_inlinemath.py
index dace6606..e4c50e9b 100644
--- a/marker/processors/llm/llm_inlinemath.py
+++ b/marker/processors/llm/llm_inlinemath.py
@@ -8,6 +8,7 @@
 from marker.processors.llm import BaseLLMComplexBlockProcessor
 
 from marker.processors.util import text_to_spans
+from marker.renderers.markdown import MarkdownRenderer
 from marker.schema import BlockTypes
 from marker.schema.blocks import Block, InlineMath
 from marker.schema.document import Document
@@ -30,58 +31,39 @@ class LLMInlineMathProcessor(BaseLLMComplexBlockProcessor):
 
     text_math_rewriting_prompt = """You are a text correction expert specializing in accurately reproducing text from images.
 You will receive an image of a text block and a set of extracted lines corresponding to the text in the image.
-Your task is to correct any errors in the extracted lines, including math, formatting, and other inaccuracies, and output the corrected lines in a JSON format.
-The number of output lines MUST match the number of input lines.  There are {input_line_count} input lines. Stay as faithful to the original text as possible.
+Your task is to correct any errors in the extracted block, including math, formatting, and other inaccuracies, and output the corrected block in html format.  Stay as faithful to the original text as possible.
 
 **Instructions:**
 
 1. Carefully examine the provided text block image .
-2. Analyze the extracted lines.
-3. For each extracted line, compare it to the corresponding line in the image.
-4. If there are no errors in any of the extracted lines, output "No corrections needed".
-5. For each extracted line, correct any errors, including:
-    * Inline math: Ensure all mathematical expressions are correctly formatted and rendered.  Surround them with <math>...</math> tags.
-    * Formatting: Maintain consistent formatting with the text block image, including spacing, indentation, subscripts/superscripts, and special characters.  Use the <i>, <b>, <sup>, <sub>, and <span> tags to format the text as needed.
+2. Analyze the text that has been extracted from the block.
+3. Compare the extracted text to the corresponding text in the image.
+4. If there are no errors in any of the extracted text, output "No corrections needed".
+5. Correct any errors in the extracted text, including:
+    * Inline math: Ensure all mathematical expressions are correctly formatted and rendered.  Surround them with <math>...</math> tags.  The math expressions should be rendered in simple, concise, KaTeX-compatible LaTeX.
+      * If a math expression is not in LaTeX format, convert it to LaTeX format, and surround it with <math>...</math> tags.
+    * Formatting: Maintain consistent formatting with the text block image, including spacing, indentation, subscripts/superscripts, and special characters.  Use the <h1>, <h2>, <h3>, <h4>, <i>, <b>, <sup>, <sub>, and <span> tags to format the text as needed.
     * Other inaccuracies:  If the image is handwritten then you may correct any spelling errors, or other discrepancies.
-6. Do not remove any formatting i.e bold, italics, math, superscripts, subscripts, etc from the extracted lines unless it is necessary to correct an error.
-7. The number of corrected lines in the output MUST equal the number of extracted lines provided in the input. Do not add or remove lines.
-8. Output the corrected lines in JSON format, as shown in the example below.  Each line should be in HTML format. Only use the math, br, a, i, b, sup, sub, and span tags.
+6. Do not remove any formatting i.e bold, italics, math, superscripts, subscripts, etc from the extracted text unless it is necessary to correct an error.
+7. Output the corrected text in html format, as shown in the example below.  Only use the h1, h2, h3, h4, p, math, br, a, i, b, sup, sub, and span tags.
 9. You absolutely cannot remove any <a href='#...'>...</a> tags, those are extremely important for references and are coming directly from the document, you MUST always preserve them.
 
 **Example:**
 
 Input:
-```
-{
- "extracted_lines": [
-  "Adversarial training (AT) <a href='#page-9-1'>[23]</a>, which aims to minimize\n",
-  "the model's risk under the worst-case perturbations, is cur-\n",
-  "rently the most effective approach for improving the robust-\n",
-  "ness of deep neural networks. For a given neural network\n",
-  "f(x, w) with parameters w, the optimization objective of\n",
-  "AT can be formulated as follows:\n"
- ]
-}
+```html
+Adversarial training (AT) <a href='#page-9-1'>[23]</a>, which aims to minimize the model's risk under the worst-case perturbations, is currently the most effective approach for improving the robustness of deep neural networks. For a given neural network f(x, w) with parameters w, the optimization objective of AT can be formulated as follows:
 ```
 
 Output:
 
-```json
-{
- "corrected_lines": [
-  "Adversarial training (AT) <a href='#page-9-1'>[23]</a>, which aims to minimize\n",
-  "the model's risk under the worst-case perturbations, is cur-\n",
-  "rently the most effective approach for improving the robust-\n",
-  "ness of deep neural networks. For a given neural network\n",
-  "<math>f(x, w)</math> with parameters <math>w</math>, the optimization objective of\n",
-  "AT can be formulated as follows:\n"
- ]
-}
+```html
+Adversarial training <i>(AT)</i> <a href='#page-9-1'>[23]</a>, which aims to minimize the model's risk under the worst-case perturbations, is currently the most effective approach for improving the robustness of deep neural networks. For a given neural network <math>f(x, w)</math> with parameters <math>w</math>, the optimization objective of AT can be formulated as follows:
 ```
 
 **Input:**
-```json
-{extracted_lines}
+```html
+{extracted_html}
 ```
 """
 
@@ -100,7 +82,7 @@ def rewrite_blocks(self, document: Document):
         detected_blocks = [
             (page, block)
             for page in document.pages
-            for block in page.contained_blocks(document, (BlockTypes.Text, BlockTypes.Caption, BlockTypes.SectionHeader, BlockTypes.Footnote))
+            for block in page.contained_blocks(document, (BlockTypes.Text, BlockTypes.Caption, BlockTypes.SectionHeader, BlockTypes.Footnote, BlockTypes.ListItem))
             if any([b.formats and "math" in b.formats for b in block.contained_blocks(document, (BlockTypes.Line,))])
         ]
 
@@ -141,6 +123,10 @@ def rewrite_blocks(self, document: Document):
 
         pbar.close()
 
+    def get_block_text(self, block: Block, document: Document) -> str:
+        html = block.render(document).html
+        return html
+
     def get_block_lines(self, block: Block, document: Document) -> Tuple[list, list]:
         text_lines = block.contained_blocks(document, (BlockTypes.Line,))
         extracted_lines = [line.formatted_text(document) for line in text_lines]
@@ -149,56 +135,30 @@ def get_block_lines(self, block: Block, document: Document) -> Tuple[list, list]
     def process_rewriting(self, document: Document, page: PageGroup, block: Block):
         SpanClass = get_block_class(BlockTypes.Span)
 
-        text_lines, extracted_lines = self.get_block_lines(block, document)
-        prompt = (self.text_math_rewriting_prompt.replace("{extracted_lines}",
-                                                         json.dumps({"extracted_lines": extracted_lines}, indent=2))
-                                                    .replace("{input_line_count}", str(len(extracted_lines)))
-                  )
+        block_text = self.get_block_text(block, document)
+        prompt = self.text_math_rewriting_prompt.replace("{extracted_html}", block_text)
 
         image = self.extract_image(document, block)
         response = self.llm_service(prompt, image, block, LLMTextSchema)
 
-        if not response or "corrected_lines" not in response:
+        if not response or "corrected_html" not in response:
             block.update_metadata(llm_error_count=1)
             return
 
-        corrected_lines = response["corrected_lines"]
-        if not corrected_lines:
+        corrected_html = response["corrected_html"]
+        if not corrected_html:
             block.update_metadata(llm_error_count=1)
             return
 
         # Block is fine
-        if "no corrections needed" in str(corrected_lines).lower():
+        if "no corrections needed" in corrected_html.lower():
             return
 
-        if len(corrected_lines) != len(extracted_lines):
+        if len(corrected_html) < len(block_text) * 0.6:
             block.update_metadata(llm_error_count=1)
             return
 
-        for text_line, corrected_text in zip(text_lines, corrected_lines):
-            text_line.structure = []
-            corrected_spans = text_to_spans(corrected_text)
-
-            for span_idx, span in enumerate(corrected_spans):
-                if span_idx == len(corrected_spans) - 1:
-                    span['content'] += "\n"
-
-                span_block = page.add_full_block(
-                    SpanClass(
-                        polygon=text_line.polygon,
-                        text=span['content'],
-                        font='Unknown',
-                        font_weight=0,
-                        font_size=0,
-                        minimum_position=0,
-                        maximum_position=0,
-                        formats=[span['type']],
-                        url=span.get('url'),
-                        page_id=text_line.page_id,
-                        text_extraction_method="gemini",
-                    )
-                )
-                text_line.structure.append(span_block.id)
+        block.html = corrected_html
 
 class LLMTextSchema(BaseModel):
-    corrected_lines: List[str]
\ No newline at end of file
+    corrected_html: str
\ No newline at end of file
diff --git a/marker/processors/llm/llm_text.py b/marker/processors/llm/llm_text.py
index cc475688..1fc3018a 100644
--- a/marker/processors/llm/llm_text.py
+++ b/marker/processors/llm/llm_text.py
@@ -33,7 +33,7 @@ class LLMTextProcessor(BaseLLMSimpleBlockProcessor):
 2. Analyze the extracted lines.
 3. For each extracted line, compare it to the corresponding line in the image.
 4. Correct any errors in the extracted line, including:
-    * Inline math: Ensure all mathematical expressions are correctly formatted and rendered.  Use the `<math>` and `</math>` tags to surround inline math properly.  Make sure the opening and closing tags appear in pairs, on the same line.
+    * Inline math: Ensure all mathematical expressions are correctly formatted and rendered.  Use the `<math>` and `</math>` tags to surround inline math properly.  Make sure the opening and closing tags appear in pairs, on the same line.  The math should be written in simple, concise, KaTeX-compatible LaTeX.
     * Formatting: Maintain consistent formatting with the text block image, including spacing, indentation, subscripts/superscripts, and special characters.  Use the `<i>`, `<b>`, `<sup>`, `<sub>`, and `<span>` tags to format the text as needed.
     * Other inaccuracies:  If the image is handwritten then you may correct any spelling errors, or other discrepancies.
 5. Do not remove any formatting i.e bold, italics, math, superscripts, subscripts, etc from the extracted lines unless it is necessary to correct an error.  The formatting 
diff --git a/marker/schema/blocks/base.py b/marker/schema/blocks/base.py
index 6bfaa9fc..bea221e3 100644
--- a/marker/schema/blocks/base.py
+++ b/marker/schema/blocks/base.py
@@ -266,3 +266,8 @@ def update_metadata(self, **kwargs):
                 setattr(self.metadata, key, metadata_attr + value)
             else:
                 raise ValueError(f"Metadata attribute {key} is not an integer")
+
+    def handle_html_output(self, document, child_blocks, parent_structure):
+        child_ref_blocks = [block for block in child_blocks if block.id.block_type == BlockTypes.Reference]
+        html = Block.assemble_html(self, document, child_ref_blocks, parent_structure)
+        return html + self.html
diff --git a/marker/schema/blocks/caption.py b/marker/schema/blocks/caption.py
index 15741388..472760ad 100644
--- a/marker/schema/blocks/caption.py
+++ b/marker/schema/blocks/caption.py
@@ -6,4 +6,11 @@ class Caption(Block):
     block_type: BlockTypes = BlockTypes.Caption
     block_description: str = "A text caption that is directly above or below an image or table. Only used for text describing the image or table.  "
     replace_output_newlines: bool = True
+    html: str | None = None
+
+    def assemble_html(self, document, child_blocks, parent_structure):
+        if self.html:
+            return super().handle_html_output(document, child_blocks, parent_structure)
+
+        return super().assemble_html(document, child_blocks, parent_structure)
 
diff --git a/marker/schema/blocks/footnote.py b/marker/schema/blocks/footnote.py
index b80c83e5..a52297c7 100644
--- a/marker/schema/blocks/footnote.py
+++ b/marker/schema/blocks/footnote.py
@@ -6,3 +6,10 @@ class Footnote(Block):
     block_type: BlockTypes = BlockTypes.Footnote
     block_description: str = "A footnote that explains a term or concept in the document."
     replace_output_newlines: bool = True
+    html: str | None = None
+
+    def assemble_html(self, document, child_blocks, parent_structure):
+        if self.html:
+            return super().handle_html_output(document, child_blocks, parent_structure)
+
+        return super().assemble_html(document, child_blocks, parent_structure)
diff --git a/marker/schema/blocks/inlinemath.py b/marker/schema/blocks/inlinemath.py
index d669406a..af63766f 100644
--- a/marker/schema/blocks/inlinemath.py
+++ b/marker/schema/blocks/inlinemath.py
@@ -8,11 +8,15 @@ class InlineMath(Block):
     blockquote: bool = False
     blockquote_level: int = 0
     block_description: str = "A text block that contains inline math.  This is not used for italic text or references - only for text that contains math."
+    html: str | None = None
 
     def assemble_html(self, document, child_blocks, parent_structure):
         if self.ignore_for_output:
             return ""
 
+        if self.html:
+            return super().handle_html_output(document, child_blocks, parent_structure)
+
         template = super().assemble_html(document, child_blocks, parent_structure)
         template = template.replace("\n", " ")
 
diff --git a/marker/schema/blocks/listitem.py b/marker/schema/blocks/listitem.py
index d8c45e6e..4700ce7e 100644
--- a/marker/schema/blocks/listitem.py
+++ b/marker/schema/blocks/listitem.py
@@ -20,6 +20,7 @@ class ListItem(Block):
     block_type: BlockTypes = BlockTypes.ListItem
     list_indent_level: int = 0
     block_description: str = "A list item that is part of a list.  This block is used to represent a single item in a list."
+    html: str | None = None
 
     def assemble_html(self, document, child_blocks, parent_structure):
         template = super().assemble_html(document, child_blocks, parent_structure)
@@ -27,6 +28,10 @@ def assemble_html(self, document, child_blocks, parent_structure):
         # Remove the first bullet character
         replace_bullets(child_blocks)
 
+        if self.html:
+            template = super().handle_html_output(document, child_blocks, parent_structure).strip()
+            template = template.replace("<li>", "").replace("</li>", "")
+
         el_attr = f" block-type='{self.block_type}'"
         if self.list_indent_level:
             return f"<ul><li{el_attr} class='list-indent-{self.list_indent_level}'>{template}</li></ul>"
diff --git a/marker/schema/blocks/sectionheader.py b/marker/schema/blocks/sectionheader.py
index 32468433..3dddc544 100644
--- a/marker/schema/blocks/sectionheader.py
+++ b/marker/schema/blocks/sectionheader.py
@@ -8,11 +8,15 @@ class SectionHeader(Block):
     block_type: BlockTypes = BlockTypes.SectionHeader
     heading_level: Optional[int] = None
     block_description: str = "The header of a section of text or other blocks."
+    html: str | None = None
 
     def assemble_html(self, document, child_blocks, parent_structure):
         if self.ignore_for_output:
             return ""
 
+        if self.html:
+            return super().handle_html_output(document, child_blocks, parent_structure)
+
         template = super().assemble_html(document, child_blocks, parent_structure)
         template = template.replace("\n", " ")
         tag = f"h{self.heading_level}" if self.heading_level else "h2"
diff --git a/marker/schema/blocks/text.py b/marker/schema/blocks/text.py
index 4b56867a..70f270bd 100644
--- a/marker/schema/blocks/text.py
+++ b/marker/schema/blocks/text.py
@@ -16,9 +16,7 @@ def assemble_html(self, document, child_blocks, parent_structure):
 
         # This happens when we used an llm processor
         if self.html:
-            child_ref_blocks = [block for block in child_blocks if block.id.block_type == BlockTypes.Reference]
-            html = super().assemble_html(document, child_ref_blocks, parent_structure)
-            return html + self.html
+            return super().handle_html_output(document, child_blocks, parent_structure)
 
         template = super().assemble_html(document, child_blocks, parent_structure)
         template = template.replace("\n", " ")
diff --git a/marker/scripts/run_streamlit_app.py b/marker/scripts/run_streamlit_app.py
index 597d7213..cb483e1e 100644
--- a/marker/scripts/run_streamlit_app.py
+++ b/marker/scripts/run_streamlit_app.py
@@ -1,9 +1,13 @@
 import subprocess
 import os
+import sys
 
 
 def streamlit_app_cli():
+    argv = sys.argv[1:]
     cur_dir = os.path.dirname(os.path.abspath(__file__))
     app_path = os.path.join(cur_dir, "streamlit_app.py")
     cmd = ["streamlit", "run", app_path, "--server.fileWatcherType", "none", "--server.headless", "true"]
+    if argv:
+        cmd += ["--"] + argv
     subprocess.run(cmd, env={**os.environ, "IN_STREAMLIT": "true"})
diff --git a/marker/scripts/streamlit_app.py b/marker/scripts/streamlit_app.py
index 4bc8acaf..2445e12b 100644
--- a/marker/scripts/streamlit_app.py
+++ b/marker/scripts/streamlit_app.py
@@ -1,8 +1,10 @@
 import os
+import sys
 os.environ["PYTORCH_ENABLE_MPS_FALLBACK"] = "1"
 os.environ["IN_STREAMLIT"] = "true"
 
 from marker.settings import settings
+from marker.config.printer import CustomClickPrinter
 from streamlit.runtime.uploaded_file_manager import UploadedFile
 
 import base64
@@ -12,6 +14,7 @@
 import string
 import tempfile
 from typing import Any, Dict
+import click
 
 import pypdfium2
 import streamlit as st
@@ -43,6 +46,29 @@
     BLOCKS_VIZ_TMPL = string.Template(f.read())
 
 
+@st.cache_data()
+def parse_args():
+    # Use to grab common cli options
+    @ConfigParser.common_options
+    def options_func():
+        pass
+
+    def extract_click_params(decorated_function):
+        if hasattr(decorated_function, '__click_params__'):
+            return decorated_function.__click_params__
+        return []
+
+    cmd = CustomClickPrinter("Marker app.")
+    extracted_params = extract_click_params(options_func)
+    cmd.params.extend(extracted_params)
+    ctx = click.Context(cmd)
+    try:
+        cmd_args = sys.argv[1:]
+        cmd.parse_args(ctx, cmd_args)
+        return ctx.params
+    except click.exceptions.ClickException as e:
+        return {"error": str(e)}
+
 @st.cache_resource()
 def load_models():
     return create_model_dict()
@@ -142,6 +168,7 @@ def block_display(image: Image, blocks: dict | None = None, dpi=96):
 col1, col2 = st.columns([.5, .5])
 
 model_dict = load_models()
+cli_options = parse_args()
 
 
 st.markdown("""
@@ -188,7 +215,7 @@ def block_display(image: Image, blocks: dict | None = None, dpi=96):
     with open(temp_pdf, 'wb') as f:
         f.write(in_file.getvalue())
     
-    cli_options = {
+    cli_options.update({
         "output_format": output_format,
         "page_range": page_range,
         "force_ocr": force_ocr,
@@ -196,7 +223,7 @@ def block_display(image: Image, blocks: dict | None = None, dpi=96):
         "output_dir": settings.DEBUG_DATA_FOLDER if debug else None,
         "use_llm": use_llm,
         "strip_existing_ocr": strip_existing_ocr
-    }
+    })
     config_parser = ConfigParser(cli_options)
     rendered = convert_pdf(
         temp_pdf,

From 65e91d19c5d3bead5d7a942724495c848d62a213 Mon Sep 17 00:00:00 2001
From: Vik Paruchuri <vik.paruchuri@gmail.com>
Date: Fri, 28 Feb 2025 11:16:44 -0500
Subject: [PATCH 40/46] Iterate on llm processors

---
 marker/builders/structure.py            | 14 ++++++++++++++
 marker/processors/llm/llm_equation.py   |  9 +++++++--
 marker/processors/llm/llm_inlinemath.py |  4 +---
 marker/processors/llm/llm_text.py       |  2 +-
 marker/services/claude.py               |  2 +-
 5 files changed, 24 insertions(+), 7 deletions(-)

diff --git a/marker/builders/structure.py b/marker/builders/structure.py
index 1396984a..32a3c84f 100644
--- a/marker/builders/structure.py
+++ b/marker/builders/structure.py
@@ -2,6 +2,7 @@
 
 from marker.builders import BaseBuilder
 from marker.schema import BlockTypes
+from marker.schema.blocks import Text
 from marker.schema.document import Document
 from marker.schema.groups import ListGroup
 from marker.schema.groups.page import PageGroup
@@ -28,6 +29,7 @@ def __call__(self, document: Document):
         for page in document.pages:
             self.group_caption_blocks(page)
             self.group_lists(page)
+            self.unmark_lists(page)
 
     def group_caption_blocks(self, page: PageGroup):
         gap_threshold_px = self.gap_threshold * page.polygon.height
@@ -110,3 +112,15 @@ def group_lists(self, page: PageGroup):
                 remove_ids.extend(block_structure)
 
         page.remove_structure_items(remove_ids)
+
+    def unmark_lists(self, page: PageGroup):
+        # If lists aren't grouped, unmark them as list items
+        for block_id in page.structure:
+            block = page.get_block(block_id)
+            if block.block_type == BlockTypes.ListItem:
+                generated_block = Text(
+                    polygon=block.polygon,
+                    page_id=block.page_id,
+                    structure=block.structure,
+                )
+                page.replace_block(block, generated_block)
diff --git a/marker/processors/llm/llm_equation.py b/marker/processors/llm/llm_equation.py
index 06b616a1..03bb2d48 100644
--- a/marker/processors/llm/llm_equation.py
+++ b/marker/processors/llm/llm_equation.py
@@ -30,7 +30,7 @@ class LLMEquationProcessor(BaseLLMSimpleBlockProcessor):
 
 Some guidelines:
 - Output valid html, where all the equations can render properly.
-- Use <math display="block"> as a block equation delimiter and <math> for inline equations.
+- Use <math display="block"> as a block equation delimiter and <math> for inline equations.  Do not use $ or $$ as delimiters.
 - Keep the LaTeX code inside the math tags simple, concise, and KaTeX compatible.
 - Enclose all equations in the correct math tags. Use multiple math tags inside the html to represent multiple equations.
 - Only use the html tags math, i, b, p, and br.
@@ -103,7 +103,12 @@ def rewrite_block(self, response: dict, prompt_data: PromptData, document: Docum
             return
 
         html_equation = response["html_equation"]
-        if len(html_equation) < len(text) * .5:
+        balanced_tags = html_equation.count("<math") == html_equation.count("</math>")
+        if not all([
+            html_equation,
+            balanced_tags,
+            len(html_equation) > len(text) * .3,
+        ]):
             block.update_metadata(llm_error_count=1)
             return
 
diff --git a/marker/processors/llm/llm_inlinemath.py b/marker/processors/llm/llm_inlinemath.py
index e4c50e9b..a0aa9e5c 100644
--- a/marker/processors/llm/llm_inlinemath.py
+++ b/marker/processors/llm/llm_inlinemath.py
@@ -40,7 +40,7 @@ class LLMInlineMathProcessor(BaseLLMComplexBlockProcessor):
 3. Compare the extracted text to the corresponding text in the image.
 4. If there are no errors in any of the extracted text, output "No corrections needed".
 5. Correct any errors in the extracted text, including:
-    * Inline math: Ensure all mathematical expressions are correctly formatted and rendered.  Surround them with <math>...</math> tags.  The math expressions should be rendered in simple, concise, KaTeX-compatible LaTeX.
+    * Inline math: Ensure all mathematical expressions are correctly formatted and rendered.  Surround them with <math>...</math> tags.  The math expressions should be rendered in simple, concise, KaTeX-compatible LaTeX.  Do not use $ or $$ as delimiters.
       * If a math expression is not in LaTeX format, convert it to LaTeX format, and surround it with <math>...</math> tags.
     * Formatting: Maintain consistent formatting with the text block image, including spacing, indentation, subscripts/superscripts, and special characters.  Use the <h1>, <h2>, <h3>, <h4>, <i>, <b>, <sup>, <sub>, and <span> tags to format the text as needed.
     * Other inaccuracies:  If the image is handwritten then you may correct any spelling errors, or other discrepancies.
@@ -133,8 +133,6 @@ def get_block_lines(self, block: Block, document: Document) -> Tuple[list, list]
         return text_lines, extracted_lines
 
     def process_rewriting(self, document: Document, page: PageGroup, block: Block):
-        SpanClass = get_block_class(BlockTypes.Span)
-
         block_text = self.get_block_text(block, document)
         prompt = self.text_math_rewriting_prompt.replace("{extracted_html}", block_text)
 
diff --git a/marker/processors/llm/llm_text.py b/marker/processors/llm/llm_text.py
index 1fc3018a..4510eece 100644
--- a/marker/processors/llm/llm_text.py
+++ b/marker/processors/llm/llm_text.py
@@ -33,7 +33,7 @@ class LLMTextProcessor(BaseLLMSimpleBlockProcessor):
 2. Analyze the extracted lines.
 3. For each extracted line, compare it to the corresponding line in the image.
 4. Correct any errors in the extracted line, including:
-    * Inline math: Ensure all mathematical expressions are correctly formatted and rendered.  Use the `<math>` and `</math>` tags to surround inline math properly.  Make sure the opening and closing tags appear in pairs, on the same line.  The math should be written in simple, concise, KaTeX-compatible LaTeX.
+    * Inline math: Ensure all mathematical expressions are correctly formatted and rendered.  Use the `<math>` and `</math>` tags to surround inline math properly.  Make sure the opening and closing tags appear in pairs, on the same line.  The math should be written in simple, concise, KaTeX-compatible LaTeX.  Do not use $ or $$ as delimiters.
     * Formatting: Maintain consistent formatting with the text block image, including spacing, indentation, subscripts/superscripts, and special characters.  Use the `<i>`, `<b>`, `<sup>`, `<sub>`, and `<span>` tags to format the text as needed.
     * Other inaccuracies:  If the image is handwritten then you may correct any spelling errors, or other discrepancies.
 5. Do not remove any formatting i.e bold, italics, math, superscripts, subscripts, etc from the extracted lines unless it is necessary to correct an error.  The formatting 
diff --git a/marker/services/claude.py b/marker/services/claude.py
index cb1055da..3dcd84b8 100644
--- a/marker/services/claude.py
+++ b/marker/services/claude.py
@@ -17,7 +17,7 @@ class ClaudeService(BaseService):
     claude_model_name: Annotated[
         str,
         "The name of the Google model to use for the service."
-    ] = "claude-3-7-sonnet-20250219"
+    ] = "claude-3-5-haiku-20241022"
     claude_api_key: Annotated[
         str,
         "The Claude API key to use for the service."

From 9f5e5f7d1483bf5a1a9bb8525ff6319025215c26 Mon Sep 17 00:00:00 2001
From: Vik Paruchuri <vik.paruchuri@gmail.com>
Date: Fri, 28 Feb 2025 13:51:12 -0500
Subject: [PATCH 41/46] Make weasyprint optional

---
 .github/workflows/benchmarks.yml |   2 +-
 .github/workflows/ci.yml         |   2 +-
 .github/workflows/publish.yml    |   2 +-
 .github/workflows/scripts.yml    |   2 +-
 README.md                        |   6 +
 marker/providers/__init__.py     |   5 +-
 marker/providers/document.py     |   3 +-
 marker/providers/epub.py         |   3 +-
 marker/providers/html.py         |   4 +-
 marker/providers/powerpoint.py   |   3 +-
 marker/providers/spreadsheet.py  |   3 +-
 poetry.lock                      | 436 +------------------------------
 pyproject.toml                   |  15 +-
 13 files changed, 38 insertions(+), 448 deletions(-)

diff --git a/.github/workflows/benchmarks.yml b/.github/workflows/benchmarks.yml
index 5b76ff15..66ae8ad4 100644
--- a/.github/workflows/benchmarks.yml
+++ b/.github/workflows/benchmarks.yml
@@ -21,7 +21,7 @@ jobs:
       - name: Install python dependencies
         run: |
           pip install poetry
-          poetry install
+          poetry install --extras "full"
       - name: Run benchmark test
         run: |
           poetry run python benchmarks/overall/overall.py --max_rows 5
diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index 84137df5..10f4e875 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -14,7 +14,7 @@ jobs:
       - name: Install python dependencies
         run: |
           pip install poetry
-          poetry install
+          poetry install --extras "full"
       - name: Run tests
         env:
           HF_TOKEN: ${{ secrets.HF_TOKEN }}
diff --git a/.github/workflows/publish.yml b/.github/workflows/publish.yml
index c10c70f2..4bd3b19c 100644
--- a/.github/workflows/publish.yml
+++ b/.github/workflows/publish.yml
@@ -15,7 +15,7 @@ jobs:
       - name: Install python dependencies
         run: |
           pip install poetry
-          poetry install
+          poetry install --extras "full"
       - name: Build package
         run: |
           poetry build
diff --git a/.github/workflows/scripts.yml b/.github/workflows/scripts.yml
index 06230580..4a6bd300 100644
--- a/.github/workflows/scripts.yml
+++ b/.github/workflows/scripts.yml
@@ -14,7 +14,7 @@ jobs:
       - name: Install python dependencies
         run: |
           pip install poetry
-          poetry install
+          poetry install --extras "full"
       - name: Download benchmark data
         run: |
           wget -O benchmark_data.zip "https://drive.google.com/uc?export=download&id=1NHrdYatR1rtqs2gPVfdvO0BAvocH8CJi"
diff --git a/README.md b/README.md
index 9a65f875..5e7b806e 100644
--- a/README.md
+++ b/README.md
@@ -66,6 +66,12 @@ Install with:
 pip install marker-pdf
 ```
 
+By default, marker will work on PDFs and images.  If you also want to use marker on XLSX, DOCX, HTML, etc, you will need to run:
+
+```shell
+pip install marker-pdf[full]
+```
+
 # Usage
 
 First, some configuration:
diff --git a/marker/providers/__init__.py b/marker/providers/__init__.py
index 85454067..b8c48373 100644
--- a/marker/providers/__init__.py
+++ b/marker/providers/__init__.py
@@ -5,8 +5,6 @@
 from pydantic import BaseModel
 
 from pdftext.schema import Reference
-from weasyprint import CSS
-from weasyprint.text.fonts import FontConfiguration
 
 from marker.logger import configure_logging
 from marker.schema.polygon import PolygonBox
@@ -75,6 +73,9 @@ def __enter__(self):
 
     @staticmethod
     def get_font_css():
+        from weasyprint import CSS
+        from weasyprint.text.fonts import FontConfiguration
+
         font_config = FontConfiguration()
         css = CSS(string=f'''
             @font-face {{
diff --git a/marker/providers/document.py b/marker/providers/document.py
index 58621d1e..00798061 100644
--- a/marker/providers/document.py
+++ b/marker/providers/document.py
@@ -7,7 +7,6 @@
 
 import mammoth
 from PIL import Image
-from weasyprint import CSS, HTML
 
 from marker.providers.pdf import PdfProvider
 
@@ -69,6 +68,8 @@ def __del__(self):
             os.remove(self.temp_pdf_path)
 
     def convert_docx_to_pdf(self, filepath: str):
+        from weasyprint import CSS, HTML
+
         with open(filepath, "rb") as docx_file:
             # we convert the docx to HTML
             result = mammoth.convert_to_html(docx_file)
diff --git a/marker/providers/epub.py b/marker/providers/epub.py
index d50ce99b..3cb25110 100644
--- a/marker/providers/epub.py
+++ b/marker/providers/epub.py
@@ -5,7 +5,6 @@
 import ebooklib
 from bs4 import BeautifulSoup
 from ebooklib import epub
-from weasyprint import CSS, HTML
 
 from marker.providers.pdf import PdfProvider
 
@@ -67,6 +66,8 @@ def __del__(self):
             os.remove(self.temp_pdf_path)
 
     def convert_epub_to_pdf(self, filepath):
+        from weasyprint import CSS, HTML
+
         ebook = epub.read_epub(filepath)
 
         styles = []
diff --git a/marker/providers/html.py b/marker/providers/html.py
index 6b24e918..ee12add5 100644
--- a/marker/providers/html.py
+++ b/marker/providers/html.py
@@ -1,8 +1,6 @@
 import os
 import tempfile
 
-from weasyprint import HTML
-
 from marker.providers.pdf import PdfProvider
 
 class HTMLProvider(PdfProvider):
@@ -26,6 +24,8 @@ def __del__(self):
             os.remove(self.temp_pdf_path)
 
     def convert_html_to_pdf(self, filepath: str):
+        from weasyprint import HTML
+
         font_css = self.get_font_css()
         HTML(filename=filepath, encoding="utf-8").write_pdf(
             self.temp_pdf_path,
diff --git a/marker/providers/powerpoint.py b/marker/providers/powerpoint.py
index 71e56d26..2bc6dcee 100644
--- a/marker/providers/powerpoint.py
+++ b/marker/providers/powerpoint.py
@@ -5,7 +5,6 @@
 
 from pptx import Presentation
 from pptx.enum.shapes import MSO_SHAPE_TYPE, PP_PLACEHOLDER
-from weasyprint import CSS, HTML
 
 from marker.providers.pdf import PdfProvider
 
@@ -63,6 +62,8 @@ def __del__(self):
             os.remove(self.temp_pdf_path)
 
     def convert_pptx_to_pdf(self, filepath):
+        from weasyprint import CSS, HTML
+
         pptx = Presentation(filepath)
 
         html_parts = []
diff --git a/marker/providers/spreadsheet.py b/marker/providers/spreadsheet.py
index 9ec3b428..c544f7d9 100644
--- a/marker/providers/spreadsheet.py
+++ b/marker/providers/spreadsheet.py
@@ -3,7 +3,6 @@
 
 from openpyxl import load_workbook
 from openpyxl.worksheet.worksheet import Worksheet
-from weasyprint import CSS, HTML
 
 from marker.providers.pdf import PdfProvider
 
@@ -52,6 +51,8 @@ def __del__(self):
             os.remove(self.temp_pdf_path)
 
     def convert_xlsx_to_pdf(self, filepath: str):
+        from weasyprint import CSS, HTML
+
         html = ""
         workbook = load_workbook(filepath)
         if workbook is not None:
diff --git a/poetry.lock b/poetry.lock
index e5f79050..eb5298b1 100644
--- a/poetry.lock
+++ b/poetry.lock
@@ -431,179 +431,6 @@ files = [
     {file = "blinker-1.9.0.tar.gz", hash = "sha256:b4ce2265a7abece45e7cc896e98dbebe6cead56bcf805a3d23136d145f5445bf"},
 ]
 
-[[package]]
-name = "brotli"
-version = "1.1.0"
-description = "Python bindings for the Brotli compression library"
-optional = false
-python-versions = "*"
-files = [
-    {file = "Brotli-1.1.0-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:e1140c64812cb9b06c922e77f1c26a75ec5e3f0fb2bf92cc8c58720dec276752"},
-    {file = "Brotli-1.1.0-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:c8fd5270e906eef71d4a8d19b7c6a43760c6abcfcc10c9101d14eb2357418de9"},
-    {file = "Brotli-1.1.0-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:1ae56aca0402a0f9a3431cddda62ad71666ca9d4dc3a10a142b9dce2e3c0cda3"},
-    {file = "Brotli-1.1.0-cp310-cp310-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:43ce1b9935bfa1ede40028054d7f48b5469cd02733a365eec8a329ffd342915d"},
-    {file = "Brotli-1.1.0-cp310-cp310-manylinux_2_5_i686.manylinux1_i686.manylinux_2_12_i686.manylinux2010_i686.whl", hash = "sha256:7c4855522edb2e6ae7fdb58e07c3ba9111e7621a8956f481c68d5d979c93032e"},
-    {file = "Brotli-1.1.0-cp310-cp310-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_12_x86_64.manylinux2010_x86_64.whl", hash = "sha256:38025d9f30cf4634f8309c6874ef871b841eb3c347e90b0851f63d1ded5212da"},
-    {file = "Brotli-1.1.0-cp310-cp310-musllinux_1_1_aarch64.whl", hash = "sha256:e6a904cb26bfefc2f0a6f240bdf5233be78cd2488900a2f846f3c3ac8489ab80"},
-    {file = "Brotli-1.1.0-cp310-cp310-musllinux_1_1_i686.whl", hash = "sha256:a37b8f0391212d29b3a91a799c8e4a2855e0576911cdfb2515487e30e322253d"},
-    {file = "Brotli-1.1.0-cp310-cp310-musllinux_1_1_ppc64le.whl", hash = "sha256:e84799f09591700a4154154cab9787452925578841a94321d5ee8fb9a9a328f0"},
-    {file = "Brotli-1.1.0-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:f66b5337fa213f1da0d9000bc8dc0cb5b896b726eefd9c6046f699b169c41b9e"},
-    {file = "Brotli-1.1.0-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:5dab0844f2cf82be357a0eb11a9087f70c5430b2c241493fc122bb6f2bb0917c"},
-    {file = "Brotli-1.1.0-cp310-cp310-musllinux_1_2_i686.whl", hash = "sha256:e4fe605b917c70283db7dfe5ada75e04561479075761a0b3866c081d035b01c1"},
-    {file = "Brotli-1.1.0-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:1e9a65b5736232e7a7f91ff3d02277f11d339bf34099a56cdab6a8b3410a02b2"},
-    {file = "Brotli-1.1.0-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:58d4b711689366d4a03ac7957ab8c28890415e267f9b6589969e74b6e42225ec"},
-    {file = "Brotli-1.1.0-cp310-cp310-win32.whl", hash = "sha256:be36e3d172dc816333f33520154d708a2657ea63762ec16b62ece02ab5e4daf2"},
-    {file = "Brotli-1.1.0-cp310-cp310-win_amd64.whl", hash = "sha256:0c6244521dda65ea562d5a69b9a26120769b7a9fb3db2fe9545935ed6735b128"},
-    {file = "Brotli-1.1.0-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:a3daabb76a78f829cafc365531c972016e4aa8d5b4bf60660ad8ecee19df7ccc"},
-    {file = "Brotli-1.1.0-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:c8146669223164fc87a7e3de9f81e9423c67a79d6b3447994dfb9c95da16e2d6"},
-    {file = "Brotli-1.1.0-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:30924eb4c57903d5a7526b08ef4a584acc22ab1ffa085faceb521521d2de32dd"},
-    {file = "Brotli-1.1.0-cp311-cp311-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:ceb64bbc6eac5a140ca649003756940f8d6a7c444a68af170b3187623b43bebf"},
-    {file = "Brotli-1.1.0-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:a469274ad18dc0e4d316eefa616d1d0c2ff9da369af19fa6f3daa4f09671fd61"},
-    {file = "Brotli-1.1.0-cp311-cp311-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:524f35912131cc2cabb00edfd8d573b07f2d9f21fa824bd3fb19725a9cf06327"},
-    {file = "Brotli-1.1.0-cp311-cp311-musllinux_1_1_aarch64.whl", hash = "sha256:5b3cc074004d968722f51e550b41a27be656ec48f8afaeeb45ebf65b561481dd"},
-    {file = "Brotli-1.1.0-cp311-cp311-musllinux_1_1_i686.whl", hash = "sha256:19c116e796420b0cee3da1ccec3b764ed2952ccfcc298b55a10e5610ad7885f9"},
-    {file = "Brotli-1.1.0-cp311-cp311-musllinux_1_1_ppc64le.whl", hash = "sha256:510b5b1bfbe20e1a7b3baf5fed9e9451873559a976c1a78eebaa3b86c57b4265"},
-    {file = "Brotli-1.1.0-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:a1fd8a29719ccce974d523580987b7f8229aeace506952fa9ce1d53a033873c8"},
-    {file = "Brotli-1.1.0-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:c247dd99d39e0338a604f8c2b3bc7061d5c2e9e2ac7ba9cc1be5a69cb6cd832f"},
-    {file = "Brotli-1.1.0-cp311-cp311-musllinux_1_2_i686.whl", hash = "sha256:1b2c248cd517c222d89e74669a4adfa5577e06ab68771a529060cf5a156e9757"},
-    {file = "Brotli-1.1.0-cp311-cp311-musllinux_1_2_ppc64le.whl", hash = "sha256:2a24c50840d89ded6c9a8fdc7b6ed3692ed4e86f1c4a4a938e1e92def92933e0"},
-    {file = "Brotli-1.1.0-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:f31859074d57b4639318523d6ffdca586ace54271a73ad23ad021acd807eb14b"},
-    {file = "Brotli-1.1.0-cp311-cp311-win32.whl", hash = "sha256:39da8adedf6942d76dc3e46653e52df937a3c4d6d18fdc94a7c29d263b1f5b50"},
-    {file = "Brotli-1.1.0-cp311-cp311-win_amd64.whl", hash = "sha256:aac0411d20e345dc0920bdec5548e438e999ff68d77564d5e9463a7ca9d3e7b1"},
-    {file = "Brotli-1.1.0-cp312-cp312-macosx_10_13_universal2.whl", hash = "sha256:32d95b80260d79926f5fab3c41701dbb818fde1c9da590e77e571eefd14abe28"},
-    {file = "Brotli-1.1.0-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:b760c65308ff1e462f65d69c12e4ae085cff3b332d894637f6273a12a482d09f"},
-    {file = "Brotli-1.1.0-cp312-cp312-macosx_10_9_universal2.whl", hash = "sha256:316cc9b17edf613ac76b1f1f305d2a748f1b976b033b049a6ecdfd5612c70409"},
-    {file = "Brotli-1.1.0-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:caf9ee9a5775f3111642d33b86237b05808dafcd6268faa492250e9b78046eb2"},
-    {file = "Brotli-1.1.0-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:70051525001750221daa10907c77830bc889cb6d865cc0b813d9db7fefc21451"},
-    {file = "Brotli-1.1.0-cp312-cp312-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:7f4bf76817c14aa98cc6697ac02f3972cb8c3da93e9ef16b9c66573a68014f91"},
-    {file = "Brotli-1.1.0-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:d0c5516f0aed654134a2fc936325cc2e642f8a0e096d075209672eb321cff408"},
-    {file = "Brotli-1.1.0-cp312-cp312-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:6c3020404e0b5eefd7c9485ccf8393cfb75ec38ce75586e046573c9dc29967a0"},
-    {file = "Brotli-1.1.0-cp312-cp312-musllinux_1_1_aarch64.whl", hash = "sha256:4ed11165dd45ce798d99a136808a794a748d5dc38511303239d4e2363c0695dc"},
-    {file = "Brotli-1.1.0-cp312-cp312-musllinux_1_1_i686.whl", hash = "sha256:4093c631e96fdd49e0377a9c167bfd75b6d0bad2ace734c6eb20b348bc3ea180"},
-    {file = "Brotli-1.1.0-cp312-cp312-musllinux_1_1_ppc64le.whl", hash = "sha256:7e4c4629ddad63006efa0ef968c8e4751c5868ff0b1c5c40f76524e894c50248"},
-    {file = "Brotli-1.1.0-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:861bf317735688269936f755fa136a99d1ed526883859f86e41a5d43c61d8966"},
-    {file = "Brotli-1.1.0-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:87a3044c3a35055527ac75e419dfa9f4f3667a1e887ee80360589eb8c90aabb9"},
-    {file = "Brotli-1.1.0-cp312-cp312-musllinux_1_2_i686.whl", hash = "sha256:c5529b34c1c9d937168297f2c1fde7ebe9ebdd5e121297ff9c043bdb2ae3d6fb"},
-    {file = "Brotli-1.1.0-cp312-cp312-musllinux_1_2_ppc64le.whl", hash = "sha256:ca63e1890ede90b2e4454f9a65135a4d387a4585ff8282bb72964fab893f2111"},
-    {file = "Brotli-1.1.0-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:e79e6520141d792237c70bcd7a3b122d00f2613769ae0cb61c52e89fd3443839"},
-    {file = "Brotli-1.1.0-cp312-cp312-win32.whl", hash = "sha256:5f4d5ea15c9382135076d2fb28dde923352fe02951e66935a9efaac8f10e81b0"},
-    {file = "Brotli-1.1.0-cp312-cp312-win_amd64.whl", hash = "sha256:906bc3a79de8c4ae5b86d3d75a8b77e44404b0f4261714306e3ad248d8ab0951"},
-    {file = "Brotli-1.1.0-cp313-cp313-macosx_10_13_universal2.whl", hash = "sha256:8bf32b98b75c13ec7cf774164172683d6e7891088f6316e54425fde1efc276d5"},
-    {file = "Brotli-1.1.0-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:7bc37c4d6b87fb1017ea28c9508b36bbcb0c3d18b4260fcdf08b200c74a6aee8"},
-    {file = "Brotli-1.1.0-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:3c0ef38c7a7014ffac184db9e04debe495d317cc9c6fb10071f7fefd93100a4f"},
-    {file = "Brotli-1.1.0-cp313-cp313-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:91d7cc2a76b5567591d12c01f019dd7afce6ba8cba6571187e21e2fc418ae648"},
-    {file = "Brotli-1.1.0-cp313-cp313-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:a93dde851926f4f2678e704fadeb39e16c35d8baebd5252c9fd94ce8ce68c4a0"},
-    {file = "Brotli-1.1.0-cp313-cp313-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:f0db75f47be8b8abc8d9e31bc7aad0547ca26f24a54e6fd10231d623f183d089"},
-    {file = "Brotli-1.1.0-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:6967ced6730aed543b8673008b5a391c3b1076d834ca438bbd70635c73775368"},
-    {file = "Brotli-1.1.0-cp313-cp313-musllinux_1_2_i686.whl", hash = "sha256:7eedaa5d036d9336c95915035fb57422054014ebdeb6f3b42eac809928e40d0c"},
-    {file = "Brotli-1.1.0-cp313-cp313-musllinux_1_2_ppc64le.whl", hash = "sha256:d487f5432bf35b60ed625d7e1b448e2dc855422e87469e3f450aa5552b0eb284"},
-    {file = "Brotli-1.1.0-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:832436e59afb93e1836081a20f324cb185836c617659b07b129141a8426973c7"},
-    {file = "Brotli-1.1.0-cp313-cp313-win32.whl", hash = "sha256:43395e90523f9c23a3d5bdf004733246fba087f2948f87ab28015f12359ca6a0"},
-    {file = "Brotli-1.1.0-cp313-cp313-win_amd64.whl", hash = "sha256:9011560a466d2eb3f5a6e4929cf4a09be405c64154e12df0dd72713f6500e32b"},
-    {file = "Brotli-1.1.0-cp36-cp36m-macosx_10_9_x86_64.whl", hash = "sha256:a090ca607cbb6a34b0391776f0cb48062081f5f60ddcce5d11838e67a01928d1"},
-    {file = "Brotli-1.1.0-cp36-cp36m-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:2de9d02f5bda03d27ede52e8cfe7b865b066fa49258cbab568720aa5be80a47d"},
-    {file = "Brotli-1.1.0-cp36-cp36m-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:2333e30a5e00fe0fe55903c8832e08ee9c3b1382aacf4db26664a16528d51b4b"},
-    {file = "Brotli-1.1.0-cp36-cp36m-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:4d4a848d1837973bf0f4b5e54e3bec977d99be36a7895c61abb659301b02c112"},
-    {file = "Brotli-1.1.0-cp36-cp36m-manylinux_2_5_x86_64.manylinux1_x86_64.whl", hash = "sha256:fdc3ff3bfccdc6b9cc7c342c03aa2400683f0cb891d46e94b64a197910dc4064"},
-    {file = "Brotli-1.1.0-cp36-cp36m-musllinux_1_1_aarch64.whl", hash = "sha256:5eeb539606f18a0b232d4ba45adccde4125592f3f636a6182b4a8a436548b914"},
-    {file = "Brotli-1.1.0-cp36-cp36m-musllinux_1_1_i686.whl", hash = "sha256:fd5f17ff8f14003595ab414e45fce13d073e0762394f957182e69035c9f3d7c2"},
-    {file = "Brotli-1.1.0-cp36-cp36m-musllinux_1_1_ppc64le.whl", hash = "sha256:069a121ac97412d1fe506da790b3e69f52254b9df4eb665cd42460c837193354"},
-    {file = "Brotli-1.1.0-cp36-cp36m-musllinux_1_1_x86_64.whl", hash = "sha256:e93dfc1a1165e385cc8239fab7c036fb2cd8093728cbd85097b284d7b99249a2"},
-    {file = "Brotli-1.1.0-cp36-cp36m-musllinux_1_2_aarch64.whl", hash = "sha256:aea440a510e14e818e67bfc4027880e2fb500c2ccb20ab21c7a7c8b5b4703d75"},
-    {file = "Brotli-1.1.0-cp36-cp36m-musllinux_1_2_i686.whl", hash = "sha256:6974f52a02321b36847cd19d1b8e381bf39939c21efd6ee2fc13a28b0d99348c"},
-    {file = "Brotli-1.1.0-cp36-cp36m-musllinux_1_2_ppc64le.whl", hash = "sha256:a7e53012d2853a07a4a79c00643832161a910674a893d296c9f1259859a289d2"},
-    {file = "Brotli-1.1.0-cp36-cp36m-musllinux_1_2_x86_64.whl", hash = "sha256:d7702622a8b40c49bffb46e1e3ba2e81268d5c04a34f460978c6b5517a34dd52"},
-    {file = "Brotli-1.1.0-cp36-cp36m-win32.whl", hash = "sha256:a599669fd7c47233438a56936988a2478685e74854088ef5293802123b5b2460"},
-    {file = "Brotli-1.1.0-cp36-cp36m-win_amd64.whl", hash = "sha256:d143fd47fad1db3d7c27a1b1d66162e855b5d50a89666af46e1679c496e8e579"},
-    {file = "Brotli-1.1.0-cp37-cp37m-macosx_10_9_x86_64.whl", hash = "sha256:11d00ed0a83fa22d29bc6b64ef636c4552ebafcef57154b4ddd132f5638fbd1c"},
-    {file = "Brotli-1.1.0-cp37-cp37m-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:f733d788519c7e3e71f0855c96618720f5d3d60c3cb829d8bbb722dddce37985"},
-    {file = "Brotli-1.1.0-cp37-cp37m-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:929811df5462e182b13920da56c6e0284af407d1de637d8e536c5cd00a7daf60"},
-    {file = "Brotli-1.1.0-cp37-cp37m-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:0b63b949ff929fbc2d6d3ce0e924c9b93c9785d877a21a1b678877ffbbc4423a"},
-    {file = "Brotli-1.1.0-cp37-cp37m-manylinux_2_5_x86_64.manylinux1_x86_64.whl", hash = "sha256:d192f0f30804e55db0d0e0a35d83a9fead0e9a359a9ed0285dbacea60cc10a84"},
-    {file = "Brotli-1.1.0-cp37-cp37m-musllinux_1_1_aarch64.whl", hash = "sha256:f296c40e23065d0d6650c4aefe7470d2a25fffda489bcc3eb66083f3ac9f6643"},
-    {file = "Brotli-1.1.0-cp37-cp37m-musllinux_1_1_i686.whl", hash = "sha256:919e32f147ae93a09fe064d77d5ebf4e35502a8df75c29fb05788528e330fe74"},
-    {file = "Brotli-1.1.0-cp37-cp37m-musllinux_1_1_ppc64le.whl", hash = "sha256:23032ae55523cc7bccb4f6a0bf368cd25ad9bcdcc1990b64a647e7bbcce9cb5b"},
-    {file = "Brotli-1.1.0-cp37-cp37m-musllinux_1_1_x86_64.whl", hash = "sha256:224e57f6eac61cc449f498cc5f0e1725ba2071a3d4f48d5d9dffba42db196438"},
-    {file = "Brotli-1.1.0-cp37-cp37m-musllinux_1_2_aarch64.whl", hash = "sha256:cb1dac1770878ade83f2ccdf7d25e494f05c9165f5246b46a621cc849341dc01"},
-    {file = "Brotli-1.1.0-cp37-cp37m-musllinux_1_2_i686.whl", hash = "sha256:3ee8a80d67a4334482d9712b8e83ca6b1d9bc7e351931252ebef5d8f7335a547"},
-    {file = "Brotli-1.1.0-cp37-cp37m-musllinux_1_2_ppc64le.whl", hash = "sha256:5e55da2c8724191e5b557f8e18943b1b4839b8efc3ef60d65985bcf6f587dd38"},
-    {file = "Brotli-1.1.0-cp37-cp37m-musllinux_1_2_x86_64.whl", hash = "sha256:d342778ef319e1026af243ed0a07c97acf3bad33b9f29e7ae6a1f68fd083e90c"},
-    {file = "Brotli-1.1.0-cp37-cp37m-win32.whl", hash = "sha256:587ca6d3cef6e4e868102672d3bd9dc9698c309ba56d41c2b9c85bbb903cdb95"},
-    {file = "Brotli-1.1.0-cp37-cp37m-win_amd64.whl", hash = "sha256:2954c1c23f81c2eaf0b0717d9380bd348578a94161a65b3a2afc62c86467dd68"},
-    {file = "Brotli-1.1.0-cp38-cp38-macosx_10_9_universal2.whl", hash = "sha256:efa8b278894b14d6da122a72fefcebc28445f2d3f880ac59d46c90f4c13be9a3"},
-    {file = "Brotli-1.1.0-cp38-cp38-macosx_10_9_x86_64.whl", hash = "sha256:03d20af184290887bdea3f0f78c4f737d126c74dc2f3ccadf07e54ceca3bf208"},
-    {file = "Brotli-1.1.0-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:6172447e1b368dcbc458925e5ddaf9113477b0ed542df258d84fa28fc45ceea7"},
-    {file = "Brotli-1.1.0-cp38-cp38-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:a743e5a28af5f70f9c080380a5f908d4d21d40e8f0e0c8901604d15cfa9ba751"},
-    {file = "Brotli-1.1.0-cp38-cp38-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:0541e747cce78e24ea12d69176f6a7ddb690e62c425e01d31cc065e69ce55b48"},
-    {file = "Brotli-1.1.0-cp38-cp38-manylinux_2_5_x86_64.manylinux1_x86_64.whl", hash = "sha256:cdbc1fc1bc0bff1cef838eafe581b55bfbffaed4ed0318b724d0b71d4d377619"},
-    {file = "Brotli-1.1.0-cp38-cp38-musllinux_1_1_aarch64.whl", hash = "sha256:890b5a14ce214389b2cc36ce82f3093f96f4cc730c1cffdbefff77a7c71f2a97"},
-    {file = "Brotli-1.1.0-cp38-cp38-musllinux_1_1_i686.whl", hash = "sha256:1ab4fbee0b2d9098c74f3057b2bc055a8bd92ccf02f65944a241b4349229185a"},
-    {file = "Brotli-1.1.0-cp38-cp38-musllinux_1_1_ppc64le.whl", hash = "sha256:141bd4d93984070e097521ed07e2575b46f817d08f9fa42b16b9b5f27b5ac088"},
-    {file = "Brotli-1.1.0-cp38-cp38-musllinux_1_1_x86_64.whl", hash = "sha256:fce1473f3ccc4187f75b4690cfc922628aed4d3dd013d047f95a9b3919a86596"},
-    {file = "Brotli-1.1.0-cp38-cp38-musllinux_1_2_aarch64.whl", hash = "sha256:d2b35ca2c7f81d173d2fadc2f4f31e88cc5f7a39ae5b6db5513cf3383b0e0ec7"},
-    {file = "Brotli-1.1.0-cp38-cp38-musllinux_1_2_i686.whl", hash = "sha256:af6fa6817889314555aede9a919612b23739395ce767fe7fcbea9a80bf140fe5"},
-    {file = "Brotli-1.1.0-cp38-cp38-musllinux_1_2_ppc64le.whl", hash = "sha256:2feb1d960f760a575dbc5ab3b1c00504b24caaf6986e2dc2b01c09c87866a943"},
-    {file = "Brotli-1.1.0-cp38-cp38-musllinux_1_2_x86_64.whl", hash = "sha256:4410f84b33374409552ac9b6903507cdb31cd30d2501fc5ca13d18f73548444a"},
-    {file = "Brotli-1.1.0-cp38-cp38-win32.whl", hash = "sha256:db85ecf4e609a48f4b29055f1e144231b90edc90af7481aa731ba2d059226b1b"},
-    {file = "Brotli-1.1.0-cp38-cp38-win_amd64.whl", hash = "sha256:3d7954194c36e304e1523f55d7042c59dc53ec20dd4e9ea9d151f1b62b4415c0"},
-    {file = "Brotli-1.1.0-cp39-cp39-macosx_10_9_universal2.whl", hash = "sha256:5fb2ce4b8045c78ebbc7b8f3c15062e435d47e7393cc57c25115cfd49883747a"},
-    {file = "Brotli-1.1.0-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:7905193081db9bfa73b1219140b3d315831cbff0d8941f22da695832f0dd188f"},
-    {file = "Brotli-1.1.0-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:a77def80806c421b4b0af06f45d65a136e7ac0bdca3c09d9e2ea4e515367c7e9"},
-    {file = "Brotli-1.1.0-cp39-cp39-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:8dadd1314583ec0bf2d1379f7008ad627cd6336625d6679cf2f8e67081b83acf"},
-    {file = "Brotli-1.1.0-cp39-cp39-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:901032ff242d479a0efa956d853d16875d42157f98951c0230f69e69f9c09bac"},
-    {file = "Brotli-1.1.0-cp39-cp39-manylinux_2_5_x86_64.manylinux1_x86_64.whl", hash = "sha256:22fc2a8549ffe699bfba2256ab2ed0421a7b8fadff114a3d201794e45a9ff578"},
-    {file = "Brotli-1.1.0-cp39-cp39-musllinux_1_1_aarch64.whl", hash = "sha256:ae15b066e5ad21366600ebec29a7ccbc86812ed267e4b28e860b8ca16a2bc474"},
-    {file = "Brotli-1.1.0-cp39-cp39-musllinux_1_1_i686.whl", hash = "sha256:949f3b7c29912693cee0afcf09acd6ebc04c57af949d9bf77d6101ebb61e388c"},
-    {file = "Brotli-1.1.0-cp39-cp39-musllinux_1_1_ppc64le.whl", hash = "sha256:89f4988c7203739d48c6f806f1e87a1d96e0806d44f0fba61dba81392c9e474d"},
-    {file = "Brotli-1.1.0-cp39-cp39-musllinux_1_1_x86_64.whl", hash = "sha256:de6551e370ef19f8de1807d0a9aa2cdfdce2e85ce88b122fe9f6b2b076837e59"},
-    {file = "Brotli-1.1.0-cp39-cp39-musllinux_1_2_aarch64.whl", hash = "sha256:0737ddb3068957cf1b054899b0883830bb1fec522ec76b1098f9b6e0f02d9419"},
-    {file = "Brotli-1.1.0-cp39-cp39-musllinux_1_2_i686.whl", hash = "sha256:4f3607b129417e111e30637af1b56f24f7a49e64763253bbc275c75fa887d4b2"},
-    {file = "Brotli-1.1.0-cp39-cp39-musllinux_1_2_ppc64le.whl", hash = "sha256:6c6e0c425f22c1c719c42670d561ad682f7bfeeef918edea971a79ac5252437f"},
-    {file = "Brotli-1.1.0-cp39-cp39-musllinux_1_2_x86_64.whl", hash = "sha256:494994f807ba0b92092a163a0a283961369a65f6cbe01e8891132b7a320e61eb"},
-    {file = "Brotli-1.1.0-cp39-cp39-win32.whl", hash = "sha256:f0d8a7a6b5983c2496e364b969f0e526647a06b075d034f3297dc66f3b360c64"},
-    {file = "Brotli-1.1.0-cp39-cp39-win_amd64.whl", hash = "sha256:cdad5b9014d83ca68c25d2e9444e28e967ef16e80f6b436918c700c117a85467"},
-    {file = "Brotli-1.1.0.tar.gz", hash = "sha256:81de08ac11bcb85841e440c13611c00b67d3bf82698314928d0b676362546724"},
-]
-
-[[package]]
-name = "brotlicffi"
-version = "1.1.0.0"
-description = "Python CFFI bindings to the Brotli library"
-optional = false
-python-versions = ">=3.7"
-files = [
-    {file = "brotlicffi-1.1.0.0-cp37-abi3-macosx_10_9_x86_64.whl", hash = "sha256:9b7ae6bd1a3f0df532b6d67ff674099a96d22bc0948955cb338488c31bfb8851"},
-    {file = "brotlicffi-1.1.0.0-cp37-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:19ffc919fa4fc6ace69286e0a23b3789b4219058313cf9b45625016bf7ff996b"},
-    {file = "brotlicffi-1.1.0.0-cp37-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:9feb210d932ffe7798ee62e6145d3a757eb6233aa9a4e7db78dd3690d7755814"},
-    {file = "brotlicffi-1.1.0.0-cp37-abi3-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:84763dbdef5dd5c24b75597a77e1b30c66604725707565188ba54bab4f114820"},
-    {file = "brotlicffi-1.1.0.0-cp37-abi3-win32.whl", hash = "sha256:1b12b50e07c3911e1efa3a8971543e7648100713d4e0971b13631cce22c587eb"},
-    {file = "brotlicffi-1.1.0.0-cp37-abi3-win_amd64.whl", hash = "sha256:994a4f0681bb6c6c3b0925530a1926b7a189d878e6e5e38fae8efa47c5d9c613"},
-    {file = "brotlicffi-1.1.0.0-pp310-pypy310_pp73-macosx_10_9_x86_64.whl", hash = "sha256:2e4aeb0bd2540cb91b069dbdd54d458da8c4334ceaf2d25df2f4af576d6766ca"},
-    {file = "brotlicffi-1.1.0.0-pp310-pypy310_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:4b7b0033b0d37bb33009fb2fef73310e432e76f688af76c156b3594389d81391"},
-    {file = "brotlicffi-1.1.0.0-pp310-pypy310_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:54a07bb2374a1eba8ebb52b6fafffa2afd3c4df85ddd38fcc0511f2bb387c2a8"},
-    {file = "brotlicffi-1.1.0.0-pp310-pypy310_pp73-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:7901a7dc4b88f1c1475de59ae9be59799db1007b7d059817948d8e4f12e24e35"},
-    {file = "brotlicffi-1.1.0.0-pp310-pypy310_pp73-win_amd64.whl", hash = "sha256:ce01c7316aebc7fce59da734286148b1d1b9455f89cf2c8a4dfce7d41db55c2d"},
-    {file = "brotlicffi-1.1.0.0-pp37-pypy37_pp73-macosx_10_9_x86_64.whl", hash = "sha256:246f1d1a90279bb6069de3de8d75a8856e073b8ff0b09dcca18ccc14cec85979"},
-    {file = "brotlicffi-1.1.0.0-pp37-pypy37_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:cc4bc5d82bc56ebd8b514fb8350cfac4627d6b0743382e46d033976a5f80fab6"},
-    {file = "brotlicffi-1.1.0.0-pp37-pypy37_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:37c26ecb14386a44b118ce36e546ce307f4810bc9598a6e6cb4f7fca725ae7e6"},
-    {file = "brotlicffi-1.1.0.0-pp37-pypy37_pp73-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:ca72968ae4eaf6470498d5c2887073f7efe3b1e7d7ec8be11a06a79cc810e990"},
-    {file = "brotlicffi-1.1.0.0-pp37-pypy37_pp73-win_amd64.whl", hash = "sha256:add0de5b9ad9e9aa293c3aa4e9deb2b61e99ad6c1634e01d01d98c03e6a354cc"},
-    {file = "brotlicffi-1.1.0.0-pp38-pypy38_pp73-macosx_10_9_x86_64.whl", hash = "sha256:9b6068e0f3769992d6b622a1cd2e7835eae3cf8d9da123d7f51ca9c1e9c333e5"},
-    {file = "brotlicffi-1.1.0.0-pp38-pypy38_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:8557a8559509b61e65083f8782329188a250102372576093c88930c875a69838"},
-    {file = "brotlicffi-1.1.0.0-pp38-pypy38_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:2a7ae37e5d79c5bdfb5b4b99f2715a6035e6c5bf538c3746abc8e26694f92f33"},
-    {file = "brotlicffi-1.1.0.0-pp38-pypy38_pp73-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:391151ec86bb1c683835980f4816272a87eaddc46bb91cbf44f62228b84d8cca"},
-    {file = "brotlicffi-1.1.0.0-pp38-pypy38_pp73-win_amd64.whl", hash = "sha256:2f3711be9290f0453de8eed5275d93d286abe26b08ab4a35d7452caa1fef532f"},
-    {file = "brotlicffi-1.1.0.0-pp39-pypy39_pp73-macosx_10_9_x86_64.whl", hash = "sha256:1a807d760763e398bbf2c6394ae9da5815901aa93ee0a37bca5efe78d4ee3171"},
-    {file = "brotlicffi-1.1.0.0-pp39-pypy39_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:fa8ca0623b26c94fccc3a1fdd895be1743b838f3917300506d04aa3346fd2a14"},
-    {file = "brotlicffi-1.1.0.0-pp39-pypy39_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:3de0cf28a53a3238b252aca9fed1593e9d36c1d116748013339f0949bfc84112"},
-    {file = "brotlicffi-1.1.0.0-pp39-pypy39_pp73-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:6be5ec0e88a4925c91f3dea2bb0013b3a2accda6f77238f76a34a1ea532a1cb0"},
-    {file = "brotlicffi-1.1.0.0-pp39-pypy39_pp73-win_amd64.whl", hash = "sha256:d9eb71bb1085d996244439154387266fd23d6ad37161f6f52f1cd41dd95a3808"},
-    {file = "brotlicffi-1.1.0.0.tar.gz", hash = "sha256:b77827a689905143f87915310b93b273ab17888fd43ef350d4832c4a71083c13"},
-]
-
-[package.dependencies]
-cffi = ">=1.0.0"
-
 [[package]]
 name = "cachetools"
 version = "5.5.2"
@@ -859,25 +686,6 @@ traitlets = ">=4"
 [package.extras]
 test = ["pytest"]
 
-[[package]]
-name = "cssselect2"
-version = "0.7.0"
-description = "CSS selectors for Python ElementTree"
-optional = false
-python-versions = ">=3.7"
-files = [
-    {file = "cssselect2-0.7.0-py3-none-any.whl", hash = "sha256:fd23a65bfd444595913f02fc71f6b286c29261e354c41d722ca7a261a49b5969"},
-    {file = "cssselect2-0.7.0.tar.gz", hash = "sha256:1ccd984dab89fc68955043aca4e1b03e0cf29cad9880f6e28e3ba7a74b14aa5a"},
-]
-
-[package.dependencies]
-tinycss2 = "*"
-webencodings = "*"
-
-[package.extras]
-doc = ["sphinx", "sphinx_rtd_theme"]
-test = ["flake8", "isort", "pytest"]
-
 [[package]]
 name = "datasets"
 version = "2.21.0"
@@ -1129,84 +937,6 @@ files = [
     {file = "filetype-1.2.0.tar.gz", hash = "sha256:66b56cd6474bf41d8c54660347d37afcc3f7d1970648de365c102ef77548aadb"},
 ]
 
-[[package]]
-name = "fonttools"
-version = "4.56.0"
-description = "Tools to manipulate font files"
-optional = false
-python-versions = ">=3.8"
-files = [
-    {file = "fonttools-4.56.0-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:331954d002dbf5e704c7f3756028e21db07097c19722569983ba4d74df014000"},
-    {file = "fonttools-4.56.0-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:8d1613abd5af2f93c05867b3a3759a56e8bf97eb79b1da76b2bc10892f96ff16"},
-    {file = "fonttools-4.56.0-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:705837eae384fe21cee5e5746fd4f4b2f06f87544fa60f60740007e0aa600311"},
-    {file = "fonttools-4.56.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:bc871904a53a9d4d908673c6faa15689874af1c7c5ac403a8e12d967ebd0c0dc"},
-    {file = "fonttools-4.56.0-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:38b947de71748bab150259ee05a775e8a0635891568e9fdb3cdd7d0e0004e62f"},
-    {file = "fonttools-4.56.0-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:86b2a1013ef7a64d2e94606632683f07712045ed86d937c11ef4dde97319c086"},
-    {file = "fonttools-4.56.0-cp310-cp310-win32.whl", hash = "sha256:133bedb9a5c6376ad43e6518b7e2cd2f866a05b1998f14842631d5feb36b5786"},
-    {file = "fonttools-4.56.0-cp310-cp310-win_amd64.whl", hash = "sha256:17f39313b649037f6c800209984a11fc256a6137cbe5487091c6c7187cae4685"},
-    {file = "fonttools-4.56.0-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:7ef04bc7827adb7532be3d14462390dd71287644516af3f1e67f1e6ff9c6d6df"},
-    {file = "fonttools-4.56.0-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:ffda9b8cd9cb8b301cae2602ec62375b59e2e2108a117746f12215145e3f786c"},
-    {file = "fonttools-4.56.0-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:e2e993e8db36306cc3f1734edc8ea67906c55f98683d6fd34c3fc5593fdbba4c"},
-    {file = "fonttools-4.56.0-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:003548eadd674175510773f73fb2060bb46adb77c94854af3e0cc5bc70260049"},
-    {file = "fonttools-4.56.0-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:bd9825822e7bb243f285013e653f6741954d8147427aaa0324a862cdbf4cbf62"},
-    {file = "fonttools-4.56.0-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:b23d30a2c0b992fb1c4f8ac9bfde44b5586d23457759b6cf9a787f1a35179ee0"},
-    {file = "fonttools-4.56.0-cp311-cp311-win32.whl", hash = "sha256:47b5e4680002ae1756d3ae3b6114e20aaee6cc5c69d1e5911f5ffffd3ee46c6b"},
-    {file = "fonttools-4.56.0-cp311-cp311-win_amd64.whl", hash = "sha256:14a3e3e6b211660db54ca1ef7006401e4a694e53ffd4553ab9bc87ead01d0f05"},
-    {file = "fonttools-4.56.0-cp312-cp312-macosx_10_13_universal2.whl", hash = "sha256:d6f195c14c01bd057bc9b4f70756b510e009c83c5ea67b25ced3e2c38e6ee6e9"},
-    {file = "fonttools-4.56.0-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:fa760e5fe8b50cbc2d71884a1eff2ed2b95a005f02dda2fa431560db0ddd927f"},
-    {file = "fonttools-4.56.0-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:d54a45d30251f1d729e69e5b675f9a08b7da413391a1227781e2a297fa37f6d2"},
-    {file = "fonttools-4.56.0-cp312-cp312-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:661a8995d11e6e4914a44ca7d52d1286e2d9b154f685a4d1f69add8418961563"},
-    {file = "fonttools-4.56.0-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:9d94449ad0a5f2a8bf5d2f8d71d65088aee48adbe45f3c5f8e00e3ad861ed81a"},
-    {file = "fonttools-4.56.0-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:f59746f7953f69cc3290ce2f971ab01056e55ddd0fb8b792c31a8acd7fee2d28"},
-    {file = "fonttools-4.56.0-cp312-cp312-win32.whl", hash = "sha256:bce60f9a977c9d3d51de475af3f3581d9b36952e1f8fc19a1f2254f1dda7ce9c"},
-    {file = "fonttools-4.56.0-cp312-cp312-win_amd64.whl", hash = "sha256:300c310bb725b2bdb4f5fc7e148e190bd69f01925c7ab437b9c0ca3e1c7cd9ba"},
-    {file = "fonttools-4.56.0-cp313-cp313-macosx_10_13_universal2.whl", hash = "sha256:f20e2c0dfab82983a90f3d00703ac0960412036153e5023eed2b4641d7d5e692"},
-    {file = "fonttools-4.56.0-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:f36a0868f47b7566237640c026c65a86d09a3d9ca5df1cd039e30a1da73098a0"},
-    {file = "fonttools-4.56.0-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:62b4c6802fa28e14dba010e75190e0e6228513573f1eeae57b11aa1a39b7e5b1"},
-    {file = "fonttools-4.56.0-cp313-cp313-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:a05d1f07eb0a7d755fbe01fee1fd255c3a4d3730130cf1bfefb682d18fd2fcea"},
-    {file = "fonttools-4.56.0-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:0073b62c3438cf0058488c002ea90489e8801d3a7af5ce5f7c05c105bee815c3"},
-    {file = "fonttools-4.56.0-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:e2cad98c94833465bcf28f51c248aaf07ca022efc6a3eba750ad9c1e0256d278"},
-    {file = "fonttools-4.56.0-cp313-cp313-win32.whl", hash = "sha256:d0cb73ccf7f6d7ca8d0bc7ea8ac0a5b84969a41c56ac3ac3422a24df2680546f"},
-    {file = "fonttools-4.56.0-cp313-cp313-win_amd64.whl", hash = "sha256:62cc1253827d1e500fde9dbe981219fea4eb000fd63402283472d38e7d8aa1c6"},
-    {file = "fonttools-4.56.0-cp38-cp38-macosx_10_9_universal2.whl", hash = "sha256:3fd3fccb7b9adaaecfa79ad51b759f2123e1aba97f857936ce044d4f029abd71"},
-    {file = "fonttools-4.56.0-cp38-cp38-macosx_10_9_x86_64.whl", hash = "sha256:193b86e9f769320bc98ffdb42accafb5d0c8c49bd62884f1c0702bc598b3f0a2"},
-    {file = "fonttools-4.56.0-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:6e81c1cc80c1d8bf071356cc3e0e25071fbba1c75afc48d41b26048980b3c771"},
-    {file = "fonttools-4.56.0-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:e9270505a19361e81eecdbc2c251ad1e1a9a9c2ad75fa022ccdee533f55535dc"},
-    {file = "fonttools-4.56.0-cp38-cp38-musllinux_1_2_aarch64.whl", hash = "sha256:53f5e9767978a4daf46f28e09dbeb7d010319924ae622f7b56174b777258e5ba"},
-    {file = "fonttools-4.56.0-cp38-cp38-musllinux_1_2_x86_64.whl", hash = "sha256:9da650cb29bc098b8cfd15ef09009c914b35c7986c8fa9f08b51108b7bc393b4"},
-    {file = "fonttools-4.56.0-cp38-cp38-win32.whl", hash = "sha256:965d0209e6dbdb9416100123b6709cb13f5232e2d52d17ed37f9df0cc31e2b35"},
-    {file = "fonttools-4.56.0-cp38-cp38-win_amd64.whl", hash = "sha256:654ac4583e2d7c62aebc6fc6a4c6736f078f50300e18aa105d87ce8925cfac31"},
-    {file = "fonttools-4.56.0-cp39-cp39-macosx_10_9_universal2.whl", hash = "sha256:ca7962e8e5fc047cc4e59389959843aafbf7445b6c08c20d883e60ced46370a5"},
-    {file = "fonttools-4.56.0-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:a1af375734018951c31c0737d04a9d5fd0a353a0253db5fbed2ccd44eac62d8c"},
-    {file = "fonttools-4.56.0-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:442ad4122468d0e47d83bc59d0e91b474593a8c813839e1872e47c7a0cb53b10"},
-    {file = "fonttools-4.56.0-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:3cf4f8d2a30b454ac682e12c61831dcb174950c406011418e739de592bbf8f76"},
-    {file = "fonttools-4.56.0-cp39-cp39-musllinux_1_2_aarch64.whl", hash = "sha256:96a4271f63a615bcb902b9f56de00ea225d6896052c49f20d0c91e9f43529a29"},
-    {file = "fonttools-4.56.0-cp39-cp39-musllinux_1_2_x86_64.whl", hash = "sha256:6c1d38642ca2dddc7ae992ef5d026e5061a84f10ff2b906be5680ab089f55bb8"},
-    {file = "fonttools-4.56.0-cp39-cp39-win32.whl", hash = "sha256:2d351275f73ebdd81dd5b09a8b8dac7a30f29a279d41e1c1192aedf1b6dced40"},
-    {file = "fonttools-4.56.0-cp39-cp39-win_amd64.whl", hash = "sha256:d6ca96d1b61a707ba01a43318c9c40aaf11a5a568d1e61146fafa6ab20890793"},
-    {file = "fonttools-4.56.0-py3-none-any.whl", hash = "sha256:1088182f68c303b50ca4dc0c82d42083d176cba37af1937e1a976a31149d4d14"},
-    {file = "fonttools-4.56.0.tar.gz", hash = "sha256:a114d1567e1a1586b7e9e7fc2ff686ca542a82769a296cef131e4c4af51e58f4"},
-]
-
-[package.dependencies]
-brotli = {version = ">=1.0.1", optional = true, markers = "platform_python_implementation == \"CPython\" and extra == \"woff\""}
-brotlicffi = {version = ">=0.8.0", optional = true, markers = "platform_python_implementation != \"CPython\" and extra == \"woff\""}
-zopfli = {version = ">=0.1.4", optional = true, markers = "extra == \"woff\""}
-
-[package.extras]
-all = ["brotli (>=1.0.1)", "brotlicffi (>=0.8.0)", "fs (>=2.2.0,<3)", "lxml (>=4.0)", "lz4 (>=1.7.4.2)", "matplotlib", "munkres", "pycairo", "scipy", "skia-pathops (>=0.5.0)", "sympy", "uharfbuzz (>=0.23.0)", "unicodedata2 (>=15.1.0)", "xattr", "zopfli (>=0.1.4)"]
-graphite = ["lz4 (>=1.7.4.2)"]
-interpolatable = ["munkres", "pycairo", "scipy"]
-lxml = ["lxml (>=4.0)"]
-pathops = ["skia-pathops (>=0.5.0)"]
-plot = ["matplotlib"]
-repacker = ["uharfbuzz (>=0.23.0)"]
-symfont = ["sympy"]
-type1 = ["xattr"]
-ufo = ["fs (>=2.2.0,<3)"]
-unicode = ["unicodedata2 (>=15.1.0)"]
-woff = ["brotli (>=1.0.1)", "brotlicffi (>=0.8.0)", "zopfli (>=0.1.4)"]
-
 [[package]]
 name = "fqdn"
 version = "1.5.1"
@@ -3133,10 +2863,10 @@ files = [
 
 [package.dependencies]
 numpy = [
-    {version = ">=1.26.0", markers = "python_version >= \"3.12\""},
-    {version = ">=1.23.5", markers = "python_version >= \"3.11\" and python_version < \"3.12\""},
     {version = ">=1.21.4", markers = "python_version >= \"3.10\" and platform_system == \"Darwin\" and python_version < \"3.11\""},
     {version = ">=1.21.2", markers = "platform_system != \"Darwin\" and python_version >= \"3.10\" and python_version < \"3.11\""},
+    {version = ">=1.23.5", markers = "python_version >= \"3.11\" and python_version < \"3.12\""},
+    {version = ">=1.26.0", markers = "python_version >= \"3.12\""},
 ]
 
 [[package]]
@@ -3228,9 +2958,9 @@ files = [
 
 [package.dependencies]
 numpy = [
-    {version = ">=1.26.0", markers = "python_version >= \"3.12\""},
-    {version = ">=1.23.2", markers = "python_version == \"3.11\""},
     {version = ">=1.22.4", markers = "python_version < \"3.11\""},
+    {version = ">=1.23.2", markers = "python_version == \"3.11\""},
+    {version = ">=1.26.0", markers = "python_version >= \"3.12\""},
 ]
 python-dateutil = ">=2.8.2"
 pytz = ">=2020.1"
@@ -3930,21 +3660,6 @@ numpy = ">=1.16.4"
 carto = ["pydeck-carto"]
 jupyter = ["ipykernel (>=5.1.2)", "ipython (>=5.8.0)", "ipywidgets (>=7,<8)", "traitlets (>=4.3.2)"]
 
-[[package]]
-name = "pydyf"
-version = "0.11.0"
-description = "A low-level PDF generator."
-optional = false
-python-versions = ">=3.8"
-files = [
-    {file = "pydyf-0.11.0-py3-none-any.whl", hash = "sha256:0aaf9e2ebbe786ec7a78ec3fbffa4cdcecde53fd6f563221d53c6bc1328848a3"},
-    {file = "pydyf-0.11.0.tar.gz", hash = "sha256:394dddf619cca9d0c55715e3c55ea121a9bf9cbc780cdc1201a2427917b86b64"},
-]
-
-[package.extras]
-doc = ["sphinx", "sphinx_rtd_theme"]
-test = ["pillow", "pytest", "ruff"]
-
 [[package]]
 name = "pyee"
 version = "12.1.1"
@@ -3998,21 +3713,6 @@ files = [
     {file = "pypdfium2-4.30.0.tar.gz", hash = "sha256:48b5b7e5566665bc1015b9d69c1ebabe21f6aee468b509531c3c8318eeee2e16"},
 ]
 
-[[package]]
-name = "pyphen"
-version = "0.17.2"
-description = "Pure Python module to hyphenate text"
-optional = false
-python-versions = ">=3.9"
-files = [
-    {file = "pyphen-0.17.2-py3-none-any.whl", hash = "sha256:3a07fb017cb2341e1d9ff31b8634efb1ae4dc4b130468c7c39dd3d32e7c3affd"},
-    {file = "pyphen-0.17.2.tar.gz", hash = "sha256:f60647a9c9b30ec6c59910097af82bc5dd2d36576b918e44148d8b07ef3b4aa3"},
-]
-
-[package.extras]
-doc = ["sphinx", "sphinx_rtd_theme"]
-test = ["pytest", "ruff"]
-
 [[package]]
 name = "pytest"
 version = "8.3.4"
@@ -5194,24 +4894,6 @@ webencodings = ">=0.4"
 doc = ["sphinx", "sphinx_rtd_theme"]
 test = ["pytest", "ruff"]
 
-[[package]]
-name = "tinyhtml5"
-version = "2.0.0"
-description = "HTML parser based on the WHATWG HTML specification"
-optional = false
-python-versions = ">=3.9"
-files = [
-    {file = "tinyhtml5-2.0.0-py3-none-any.whl", hash = "sha256:13683277c5b176d070f82d099d977194b7a1e26815b016114f581a74bbfbf47e"},
-    {file = "tinyhtml5-2.0.0.tar.gz", hash = "sha256:086f998833da24c300c414d9fe81d9b368fd04cb9d2596a008421cbc705fcfcc"},
-]
-
-[package.dependencies]
-webencodings = ">=0.5.1"
-
-[package.extras]
-doc = ["sphinx", "sphinx_rtd_theme"]
-test = ["pytest", "ruff"]
-
 [[package]]
 name = "tokenizers"
 version = "0.21.0"
@@ -5632,31 +5314,6 @@ files = [
     {file = "wcwidth-0.2.13.tar.gz", hash = "sha256:72ea0c06399eb286d978fdedb6923a9eb47e1c486ce63e9b4e64fc18303972b5"},
 ]
 
-[[package]]
-name = "weasyprint"
-version = "63.1"
-description = "The Awesome Document Factory"
-optional = false
-python-versions = ">=3.9"
-files = [
-    {file = "weasyprint-63.1-py3-none-any.whl", hash = "sha256:9d0319fe3ba553c9a77dc43a2d35b64a70c2b8809ad55a139a214803fde62bce"},
-    {file = "weasyprint-63.1.tar.gz", hash = "sha256:cb424e63e8dd3f14195bfe5f203527646aa40a2f00ac819f9d39b8304cec0044"},
-]
-
-[package.dependencies]
-cffi = ">=0.6"
-cssselect2 = ">=0.1"
-fonttools = {version = ">=4.0.0", extras = ["woff"]}
-Pillow = ">=9.1.0"
-pydyf = ">=0.11.0"
-Pyphen = ">=0.9.1"
-tinycss2 = ">=1.4.0"
-tinyhtml5 = ">=2.0.0b1"
-
-[package.extras]
-doc = ["sphinx", "sphinx_rtd_theme"]
-test = ["pytest", "ruff"]
-
 [[package]]
 name = "webcolors"
 version = "24.11.1"
@@ -6023,90 +5680,7 @@ idna = ">=2.0"
 multidict = ">=4.0"
 propcache = ">=0.2.0"
 
-[[package]]
-name = "zopfli"
-version = "0.2.3.post1"
-description = "Zopfli module for python"
-optional = false
-python-versions = ">=3.8"
-files = [
-    {file = "zopfli-0.2.3.post1-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:e0137dd64a493ba6a4be37405cfd6febe650a98cc1e9dca8f6b8c63b1db11b41"},
-    {file = "zopfli-0.2.3.post1-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:aa588b21044f8a74e423d8c8a4c7fc9988501878aacced793467010039c50734"},
-    {file = "zopfli-0.2.3.post1-cp310-cp310-manylinux_2_12_i686.manylinux2010_i686.whl", hash = "sha256:9f4a7ec2770e6af05f5a02733fd3900f30a9cd58e5d6d3727e14c5bcd6e7d587"},
-    {file = "zopfli-0.2.3.post1-cp310-cp310-manylinux_2_12_x86_64.manylinux2010_x86_64.whl", hash = "sha256:f7d69c1a7168ad0e9cb864e8663acb232986a0c9c9cb9801f56bf6214f53a54d"},
-    {file = "zopfli-0.2.3.post1-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:6c2d2bc8129707e34c51f9352c4636ca313b52350bbb7e04637c46c1818a2a70"},
-    {file = "zopfli-0.2.3.post1-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:39e576f93576c5c223b41d9c780bbb91fd6db4babf3223d2a4fe7bf568e2b5a8"},
-    {file = "zopfli-0.2.3.post1-cp310-cp310-musllinux_1_2_i686.whl", hash = "sha256:cbe6df25807227519debd1a57ab236f5f6bad441500e85b13903e51f93a43214"},
-    {file = "zopfli-0.2.3.post1-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:7cce242b5df12b2b172489daf19c32e5577dd2fac659eb4b17f6a6efb446fd5c"},
-    {file = "zopfli-0.2.3.post1-cp310-cp310-win32.whl", hash = "sha256:f815fcc2b2a457977724bad97fb4854022980f51ce7b136925e336b530545ae1"},
-    {file = "zopfli-0.2.3.post1-cp310-cp310-win_amd64.whl", hash = "sha256:0cc20b02a9531559945324c38302fd4ba763311632d0ec8a1a0aa9c10ea363e6"},
-    {file = "zopfli-0.2.3.post1-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:518f1f4ed35dd69ce06b552f84e6d081f07c552b4c661c5312d950a0b764a58a"},
-    {file = "zopfli-0.2.3.post1-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:615a8ac9dda265e9cc38b2a76c3142e4a9f30fea4a79c85f670850783bc6feb4"},
-    {file = "zopfli-0.2.3.post1-cp311-cp311-manylinux_2_12_i686.manylinux2010_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:a82fc2dbebe6eb908b9c665e71496f8525c1bc4d2e3a7a7722ef2b128b6227c8"},
-    {file = "zopfli-0.2.3.post1-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:37d011e92f7b9622742c905fdbed9920a1d0361df84142807ea2a528419dea7f"},
-    {file = "zopfli-0.2.3.post1-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:e63d558847166543c2c9789e6f985400a520b7eacc4b99181668b2c3aeadd352"},
-    {file = "zopfli-0.2.3.post1-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:60db20f06c3d4c5934b16cfa62a2cc5c3f0686bffe0071ed7804d3c31ab1a04e"},
-    {file = "zopfli-0.2.3.post1-cp311-cp311-musllinux_1_2_i686.whl", hash = "sha256:716cdbfc57bfd3d3e31a58e6246e8190e6849b7dbb7c4ce39ef8bbf0edb8f6d5"},
-    {file = "zopfli-0.2.3.post1-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:3a89277ed5f8c0fb2d0b46d669aa0633123aa7381f1f6118c12f15e0fb48f8ca"},
-    {file = "zopfli-0.2.3.post1-cp311-cp311-win32.whl", hash = "sha256:75a26a2307b10745a83b660c404416e984ee6fca515ec7f0765f69af3ce08072"},
-    {file = "zopfli-0.2.3.post1-cp311-cp311-win_amd64.whl", hash = "sha256:81c341d9bb87a6dbbb0d45d6e272aca80c7c97b4b210f9b6e233bf8b87242f29"},
-    {file = "zopfli-0.2.3.post1-cp312-cp312-macosx_10_13_universal2.whl", hash = "sha256:3f0197b6aa6eb3086ae9e66d6dd86c4d502b6c68b0ec490496348ae8c05ecaef"},
-    {file = "zopfli-0.2.3.post1-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:5fcfc0dc2761e4fcc15ad5d273b4d58c2e8e059d3214a7390d4d3c8e2aee644e"},
-    {file = "zopfli-0.2.3.post1-cp312-cp312-manylinux_2_12_i686.manylinux2010_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:cac2b37ab21c2b36a10b685b1893ebd6b0f83ae26004838ac817680881576567"},
-    {file = "zopfli-0.2.3.post1-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:8d5ab297d660b75c159190ce6d73035502310e40fd35170aed7d1a1aea7ddd65"},
-    {file = "zopfli-0.2.3.post1-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:9ba214f4f45bec195ee8559651154d3ac2932470b9d91c5715fc29c013349f8c"},
-    {file = "zopfli-0.2.3.post1-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:c1e0ed5d84ffa2d677cc9582fc01e61dab2e7ef8b8996e055f0a76167b1b94df"},
-    {file = "zopfli-0.2.3.post1-cp312-cp312-musllinux_1_2_i686.whl", hash = "sha256:bfa1eb759e07d8b7aa7a310a2bc535e127ee70addf90dc8d4b946b593c3e51a8"},
-    {file = "zopfli-0.2.3.post1-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:cd2c002f160502608dcc822ed2441a0f4509c52e86fcfd1a09e937278ed1ca14"},
-    {file = "zopfli-0.2.3.post1-cp312-cp312-win32.whl", hash = "sha256:7be5cc6732eb7b4df17305d8a7b293223f934a31783a874a01164703bc1be6cd"},
-    {file = "zopfli-0.2.3.post1-cp312-cp312-win_amd64.whl", hash = "sha256:4e50ffac74842c1c1018b9b73875a0d0a877c066ab06bf7cccbaa84af97e754f"},
-    {file = "zopfli-0.2.3.post1-cp313-cp313-macosx_10_13_universal2.whl", hash = "sha256:ecb7572df5372abce8073df078207d9d1749f20b8b136089916a4a0868d56051"},
-    {file = "zopfli-0.2.3.post1-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:a1cf720896d2ce998bc8e051d4b4ce0d8bec007aab6243102e8e1d22a0b2fb3f"},
-    {file = "zopfli-0.2.3.post1-cp313-cp313-manylinux_2_12_i686.manylinux2010_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:5aad740b4d4fcbaaae4887823925166ffd062db3b248b3f432198fc287381d1a"},
-    {file = "zopfli-0.2.3.post1-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:6617fb10f9e4393b331941861d73afb119cd847e88e4974bdbe8068ceef3f73f"},
-    {file = "zopfli-0.2.3.post1-cp313-cp313-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:a53b18797cdef27e019db595d66c4b077325afe2fd62145953275f53d84ce40c"},
-    {file = "zopfli-0.2.3.post1-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:b78008a69300d929ca2efeffec951b64a312e9a811e265ea4a907ab546d79fa6"},
-    {file = "zopfli-0.2.3.post1-cp313-cp313-musllinux_1_2_i686.whl", hash = "sha256:0aa5f90d6298bda02a95bc8dc8c3c19004d5a4e44bda00b67ca7431d857b4b54"},
-    {file = "zopfli-0.2.3.post1-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:2768c877f76c8a0e7519b1c86c93757f3c01492ddde55751e9988afb7eff64e1"},
-    {file = "zopfli-0.2.3.post1-cp313-cp313-win32.whl", hash = "sha256:71390dbd3fbf6ebea9a5d85ffed8c26ee1453ee09248e9b88486e30e0397b775"},
-    {file = "zopfli-0.2.3.post1-cp313-cp313-win_amd64.whl", hash = "sha256:a86eb88e06bd87e1fff31dac878965c26b0c26db59ddcf78bb0379a954b120de"},
-    {file = "zopfli-0.2.3.post1-cp38-cp38-macosx_10_9_universal2.whl", hash = "sha256:3827170de28faf144992d3d4dcf8f3998fe3c8a6a6f4a08f1d42c2ec6119d2bb"},
-    {file = "zopfli-0.2.3.post1-cp38-cp38-macosx_10_9_x86_64.whl", hash = "sha256:b0ec13f352ea5ae0fc91f98a48540512eed0767d0ec4f7f3cb92d92797983d18"},
-    {file = "zopfli-0.2.3.post1-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:5f272186e03ad55e7af09ab78055535c201b1a0bcc2944edb1768298d9c483a4"},
-    {file = "zopfli-0.2.3.post1-cp38-cp38-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:29ea74e72ffa6e291b8c6f2504ce6c146b4fe990c724c1450eb8e4c27fd31431"},
-    {file = "zopfli-0.2.3.post1-cp38-cp38-manylinux_2_5_x86_64.manylinux1_x86_64.whl", hash = "sha256:eb45a34f23da4f8bc712b6376ca5396914b0b7c09adbb001dad964eb7f3132f8"},
-    {file = "zopfli-0.2.3.post1-cp38-cp38-musllinux_1_2_aarch64.whl", hash = "sha256:6482db9876c68faac2d20a96b566ffbf65ddaadd97b222e4e73641f4f8722fc4"},
-    {file = "zopfli-0.2.3.post1-cp38-cp38-musllinux_1_2_i686.whl", hash = "sha256:95a260cafd56b8fffa679918937401c80bb38e1681c448b988022e4c3610965d"},
-    {file = "zopfli-0.2.3.post1-cp38-cp38-musllinux_1_2_x86_64.whl", hash = "sha256:676919fba7311125244eb0c4393679ac5fe856e5864a15d122bd815205369fa0"},
-    {file = "zopfli-0.2.3.post1-cp38-cp38-win32.whl", hash = "sha256:b9026a21b6d41eb0e2e63f5bc1242c3fcc43ecb770963cda99a4307863dac12e"},
-    {file = "zopfli-0.2.3.post1-cp38-cp38-win_amd64.whl", hash = "sha256:3c163911f8bad94b3e1db0a572e7c28ba681a0c91d0002ea1e4fa9264c21ef17"},
-    {file = "zopfli-0.2.3.post1-cp39-cp39-macosx_10_9_universal2.whl", hash = "sha256:b05296e8bc88c92e2b21e0a9bae4740c1551ee613c1d93a51fd28a7a0b2b6fbb"},
-    {file = "zopfli-0.2.3.post1-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:f12000a6accdd4bf0a3fa6eaa1b1c7a7bc80af0a2edf3f89d770d3dcce1d0e22"},
-    {file = "zopfli-0.2.3.post1-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:a241a68581d34d67b40c425cce3d1fd211c092f99d9250947824ccba9f491949"},
-    {file = "zopfli-0.2.3.post1-cp39-cp39-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:3657e416ffb8f31d9d3424af12122bb251befae109f2e271d87d825c92fc5b7b"},
-    {file = "zopfli-0.2.3.post1-cp39-cp39-manylinux_2_5_x86_64.manylinux1_x86_64.whl", hash = "sha256:4915a41375bdee4db749ecd07d985a0486eb688a6619f713b7bf6fbfd145e960"},
-    {file = "zopfli-0.2.3.post1-cp39-cp39-musllinux_1_2_aarch64.whl", hash = "sha256:bbe429fc50686bb2a2608a30843e36fbaa123462a5284f136c7d9e0145220bfd"},
-    {file = "zopfli-0.2.3.post1-cp39-cp39-musllinux_1_2_i686.whl", hash = "sha256:2345e713260a350bea0b01a816a469ea356bc2d63d009a0d777691ecbbcf7493"},
-    {file = "zopfli-0.2.3.post1-cp39-cp39-musllinux_1_2_x86_64.whl", hash = "sha256:fc39f5c27f962ec8660d8d20c24762431131b5d8c672b44b0a54cf2b5bcde9b9"},
-    {file = "zopfli-0.2.3.post1-cp39-cp39-win32.whl", hash = "sha256:9a6aec38a989bad7ddd1ef53f1265699e49e294d08231b5313d61293f3cd6237"},
-    {file = "zopfli-0.2.3.post1-cp39-cp39-win_amd64.whl", hash = "sha256:b3df42f52502438ee973042cc551877d24619fa1cd38ef7b7e9ac74200daca8b"},
-    {file = "zopfli-0.2.3.post1-pp310-pypy310_pp73-macosx_10_15_x86_64.whl", hash = "sha256:4c1226a7e2c7105ac31503a9bb97454743f55d88164d6d46bc138051b77f609b"},
-    {file = "zopfli-0.2.3.post1-pp310-pypy310_pp73-manylinux_2_12_i686.manylinux2010_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:48dba9251060289101343110ab47c0756f66f809bb4d1ddbb6d5c7e7752115c5"},
-    {file = "zopfli-0.2.3.post1-pp310-pypy310_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:89899641d4de97dbad8e0cde690040d078b6aea04066dacaab98e0b5a23573f2"},
-    {file = "zopfli-0.2.3.post1-pp310-pypy310_pp73-win_amd64.whl", hash = "sha256:3654bfc927bc478b1c3f3ff5056ed7b20a1a37fa108ca503256d0a699c03bbb1"},
-    {file = "zopfli-0.2.3.post1-pp38-pypy38_pp73-macosx_10_9_x86_64.whl", hash = "sha256:c4278d1873ce6e803e5d4f8d702fd3026bd67fca744aa98881324d1157ddf748"},
-    {file = "zopfli-0.2.3.post1-pp38-pypy38_pp73-win_amd64.whl", hash = "sha256:1d8cc06605519e82b16df090e17cb3990d1158861b2872c3117f1168777b81e4"},
-    {file = "zopfli-0.2.3.post1-pp39-pypy39_pp73-macosx_10_15_x86_64.whl", hash = "sha256:1f990634fd5c5c8ced8edddd8bd45fab565123b4194d6841e01811292650acae"},
-    {file = "zopfli-0.2.3.post1-pp39-pypy39_pp73-manylinux_2_12_i686.manylinux2010_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:91a2327a4d7e77471fa4fbb26991c6de4a738c6fc6a33e09bb25f56a870a4b7b"},
-    {file = "zopfli-0.2.3.post1-pp39-pypy39_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:8fbe5bcf10d01aab3513550f284c09fef32f342b36f56bfae2120a9c4d12c130"},
-    {file = "zopfli-0.2.3.post1-pp39-pypy39_pp73-win_amd64.whl", hash = "sha256:34a99592f3d9eb6f737616b5bd74b48a589fdb3cb59a01a50d636ea81d6af272"},
-    {file = "zopfli-0.2.3.post1.tar.gz", hash = "sha256:96484dc0f48be1c5d7ae9f38ed1ce41e3675fd506b27c11a6607f14b49101e99"},
-]
-
-[package.extras]
-test = ["pytest"]
-
 [metadata]
 lock-version = "2.0"
 python-versions = "^3.10"
-content-hash = "12c3ac5c7dfd66a41559c675147d24de82ab83feceb1a0fc150772d1fc8864c0"
+content-hash = "70d4b1cc842480888768dbad78c97cdedeec150ccb3433e46c030327f2eb7819"
diff --git a/pyproject.toml b/pyproject.toml
index 5b88f2a9..03481ab0 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -36,11 +36,13 @@ filetype = "^1.2.0"
 scikit-learn = "^1.6.1"
 google-genai = "^1.0.0"
 anthropic = "^0.46.0"
-mammoth = "^1.9.0"
-weasyprint = "^63.1"
-openpyxl = "^3.1.5"
-python-pptx = "^1.0.2"
-ebooklib = "^0.18"
+
+# Optional dependencies for documents
+mammoth = {version = "^1.9.0", optional = true}
+openpyxl = {version = "^3.1.5", optional = true}
+python-pptx = {version = "^1.0.2", optional = true}
+ebooklib = {version = "^0.18", optional = true}
+weasyprint = {version = "^63.1", optional = true}
 
 [tool.poetry.group.dev.dependencies]
 jupyter = "^1.0.0"
@@ -58,6 +60,9 @@ tabulate = "^0.9.0"
 latex2mathml = "^3.77.0"
 playwright = "^1.49.1"
 
+[tool.poetry.extras]
+full = ["mammoth", "openpyxl", "python-pptx", "ebooklib", "weasyprint"]
+
 [tool.poetry.scripts]
 marker = "marker.scripts.convert:convert_cli"
 marker_single = "marker.scripts.convert_single:convert_single_cli"

From ab354b0d132db4882096eef458f231c503fdf70f Mon Sep 17 00:00:00 2001
From: Vik Paruchuri <vik.paruchuri@gmail.com>
Date: Fri, 28 Feb 2025 15:16:33 -0500
Subject: [PATCH 42/46] Ensure math can wrap properly

---
 marker/processors/llm/llm_equation.py   |  32 +-
 marker/processors/llm/llm_inlinemath.py |  20 +-
 marker/processors/llm/llm_text.py       |  16 +-
 marker/renderers/markdown.py            |   9 +-
 marker/services/claude.py               |   2 +-
 poetry.lock                             | 485 ++++++++++++++++++++++--
 6 files changed, 502 insertions(+), 62 deletions(-)

diff --git a/marker/processors/llm/llm_equation.py b/marker/processors/llm/llm_equation.py
index 03bb2d48..73a62256 100644
--- a/marker/processors/llm/llm_equation.py
+++ b/marker/processors/llm/llm_equation.py
@@ -26,7 +26,7 @@ class LLMEquationProcessor(BaseLLMSimpleBlockProcessor):
         "The prompt to use for generating LaTeX from equations.",
         "Default is a string containing the Gemini prompt."
     ] = r"""You're an expert mathematician who is good at writing LaTeX code and html for equations.
-You'll receive an image of a math block that may contain one or more equations. Your job is to write html that represents the content of the image, with the equations in LaTeX format, and fenced by delimiters.
+You'll receive an image of a math block, along with the text extracted from the block.  It may contain one or more equations. Your job is to write html that represents the content of the image, with the equations in LaTeX format.
 
 Some guidelines:
 - Output valid html, where all the equations can render properly.
@@ -35,26 +35,31 @@ class LLMEquationProcessor(BaseLLMSimpleBlockProcessor):
 - Enclose all equations in the correct math tags. Use multiple math tags inside the html to represent multiple equations.
 - Only use the html tags math, i, b, p, and br.
 - Make sure to include all the equations in the image in the html output.
+- Make sure to include other text in the image in the correct positions along with the equations.
 
 **Instructions:**
 1. Carefully examine the provided image.
 2. Analyze the existing html, which may include LaTeX code.
-3. If the html and LaTeX are correct, write "No corrections needed."
-4. If the html and LaTeX are incorrect, generate the corrected html.
-5. Output only the corrected html or "No corrections needed."
+3. Write a short analysis of how the html should be corrected to represent the image.
+4. If the html and LaTeX are correct, write "No corrections needed."
+5. If the html and LaTeX are incorrect, generate the corrected html.
+6. Output only the analysis, then the corrected html or "No corrections needed."
 **Example:**
 Input:
 ```html
-Equation 1: 
-<math display="block">x2 + y2 = z2</math>
-Equation 2:
-<math display="block">\frac{ab \cdot x^5 + x^2 + 2 \cdot x + 123}{t}</math>
+The following equation illustrates the Pythagorean theorem:
+x2 + y2 = z2
+
+And this equation is a bit more complex:
+(ab * x5 + x2 + 2 * x + 123)/t
 ```
 Output:
+analysis: The equations are not formatted as LaTeX, or enclosed in math tags.
 ```html
-<p>Equation 1:</p> 
+<p>The following equation illustrates the Pythagorean theorem:</p> 
 <math display="block">x^{2} + y^{2} = z^{2}</math>
-<p>Equation 2:</p>
+
+<p>And this equation is a bit more complex, and contains <math>ab \cdot x^{5}</math>:</p>
 <math display="block">\frac{ab \cdot x^{5} + x^{2} + 2 \cdot x + 123}{t}</math>
 ```
 **Input:**
@@ -98,11 +103,11 @@ def rewrite_block(self, response: dict, prompt_data: PromptData, document: Docum
         block = prompt_data["block"]
         text = block.html if block.html else block.raw_text(document)
 
-        if not response or "html_equation" not in response:
+        if not response or "corrected_equation" not in response:
             block.update_metadata(llm_error_count=1)
             return
 
-        html_equation = response["html_equation"]
+        html_equation = response["corrected_equation"]
         balanced_tags = html_equation.count("<math") == html_equation.count("</math>")
         if not all([
             html_equation,
@@ -115,4 +120,5 @@ def rewrite_block(self, response: dict, prompt_data: PromptData, document: Docum
         block.html = html_equation
 
 class EquationSchema(BaseModel):
-    html_equation: str
+    analysis: str
+    corrected_equation: str
diff --git a/marker/processors/llm/llm_inlinemath.py b/marker/processors/llm/llm_inlinemath.py
index a0aa9e5c..0ab43907 100644
--- a/marker/processors/llm/llm_inlinemath.py
+++ b/marker/processors/llm/llm_inlinemath.py
@@ -7,13 +7,10 @@
 
 from marker.processors.llm import BaseLLMComplexBlockProcessor
 
-from marker.processors.util import text_to_spans
-from marker.renderers.markdown import MarkdownRenderer
 from marker.schema import BlockTypes
 from marker.schema.blocks import Block, InlineMath
 from marker.schema.document import Document
 from marker.schema.groups import PageGroup
-from marker.schema.registry import get_block_class
 
 
 class LLMInlineMathProcessor(BaseLLMComplexBlockProcessor):
@@ -38,25 +35,29 @@ class LLMInlineMathProcessor(BaseLLMComplexBlockProcessor):
 1. Carefully examine the provided text block image .
 2. Analyze the text that has been extracted from the block.
 3. Compare the extracted text to the corresponding text in the image.
-4. If there are no errors in any of the extracted text, output "No corrections needed".
-5. Correct any errors in the extracted text, including:
+4. Write a short analysis of the text block, including any errors you see in the extracted text.
+5. If there are no errors in any of the extracted text, output "No corrections needed".
+6. Correct any errors in the extracted text, including:
     * Inline math: Ensure all mathematical expressions are correctly formatted and rendered.  Surround them with <math>...</math> tags.  The math expressions should be rendered in simple, concise, KaTeX-compatible LaTeX.  Do not use $ or $$ as delimiters.
       * If a math expression is not in LaTeX format, convert it to LaTeX format, and surround it with <math>...</math> tags.
     * Formatting: Maintain consistent formatting with the text block image, including spacing, indentation, subscripts/superscripts, and special characters.  Use the <h1>, <h2>, <h3>, <h4>, <i>, <b>, <sup>, <sub>, and <span> tags to format the text as needed.
     * Other inaccuracies:  If the image is handwritten then you may correct any spelling errors, or other discrepancies.
-6. Do not remove any formatting i.e bold, italics, math, superscripts, subscripts, etc from the extracted text unless it is necessary to correct an error.
-7. Output the corrected text in html format, as shown in the example below.  Only use the h1, h2, h3, h4, p, math, br, a, i, b, sup, sub, and span tags.
+    * Ensure lines wrap properly, and that newlines are not in the middle of sentences.
+7. Do not remove any formatting i.e bold, italics, math, superscripts, subscripts, etc from the extracted text unless it is necessary to correct an error.
+8. Output the corrected text in html format, as shown in the example below.  Only use the h1, h2, h3, h4, p, math, br, a, i, b, sup, sub, and span tags.
 9. You absolutely cannot remove any <a href='#...'>...</a> tags, those are extremely important for references and are coming directly from the document, you MUST always preserve them.
 
 **Example:**
 
 Input:
 ```html
-Adversarial training (AT) <a href='#page-9-1'>[23]</a>, which aims to minimize the model's risk under the worst-case perturbations, is currently the most effective approach for improving the robustness of deep neural networks. For a given neural network f(x, w) with parameters w, the optimization objective of AT can be formulated as follows:
+Adversarial training (AT) <a href='#page-9-1'>[23]</a>, which aims to minimize the model's risk under the worst-case perturbations, 
+is currently the most effective approach for improving the robustness of deep neural networks. For a given neural network f(x, w) 
+with parameters w, the optimization objective of AT can be formulated as follows:
 ```
 
 Output:
-
+analysis: The inline math is not in LaTeX format and is not surrounded by <math>...</math> tags.
 ```html
 Adversarial training <i>(AT)</i> <a href='#page-9-1'>[23]</a>, which aims to minimize the model's risk under the worst-case perturbations, is currently the most effective approach for improving the robustness of deep neural networks. For a given neural network <math>f(x, w)</math> with parameters <math>w</math>, the optimization objective of AT can be formulated as follows:
 ```
@@ -159,4 +160,5 @@ def process_rewriting(self, document: Document, page: PageGroup, block: Block):
         block.html = corrected_html
 
 class LLMTextSchema(BaseModel):
+    analysis: str
     corrected_html: str
\ No newline at end of file
diff --git a/marker/processors/llm/llm_text.py b/marker/processors/llm/llm_text.py
index 4510eece..a7d85ccf 100644
--- a/marker/processors/llm/llm_text.py
+++ b/marker/processors/llm/llm_text.py
@@ -31,15 +31,16 @@ class LLMTextProcessor(BaseLLMSimpleBlockProcessor):
 
 1. Carefully examine the provided text block image .
 2. Analyze the extracted lines.
-3. For each extracted line, compare it to the corresponding line in the image.
-4. Correct any errors in the extracted line, including:
+3. Write a short analysis comparing the extracted lines to the image.
+4. For each extracted line, compare it to the corresponding line in the image.
+5. Correct any errors in the extracted line, including:
     * Inline math: Ensure all mathematical expressions are correctly formatted and rendered.  Use the `<math>` and `</math>` tags to surround inline math properly.  Make sure the opening and closing tags appear in pairs, on the same line.  The math should be written in simple, concise, KaTeX-compatible LaTeX.  Do not use $ or $$ as delimiters.
     * Formatting: Maintain consistent formatting with the text block image, including spacing, indentation, subscripts/superscripts, and special characters.  Use the `<i>`, `<b>`, `<sup>`, `<sub>`, and `<span>` tags to format the text as needed.
     * Other inaccuracies:  If the image is handwritten then you may correct any spelling errors, or other discrepancies.
-5. Do not remove any formatting i.e bold, italics, math, superscripts, subscripts, etc from the extracted lines unless it is necessary to correct an error.  The formatting 
-6. The number of corrected lines in the output MUST equal the number of extracted lines provided in the input. Do not add or remove lines.  There are exactly {line_count} input lines.
-7. Output the corrected lines in JSON format, as shown in the example below.  Each line should be in HTML format. Only use the math, br, a, i, b, sup, sub, and span tags.
-8. You absolutely cannot remove any <a href='#...'>...</a> tags, those are extremely important for references and are coming directly from the document, you MUST always preserve them.
+6. Do not remove any formatting i.e bold, italics, math, superscripts, subscripts, etc from the extracted lines unless it is necessary to correct an error.  The formatting 
+7. The number of corrected lines in the output MUST equal the number of extracted lines provided in the input. Do not add or remove lines.  There are exactly {line_count} input lines.
+8. Output the corrected lines in JSON format, as shown in the example below.  Each line should be in HTML format. Only use the math, br, a, i, b, sup, sub, and span tags.
+9. You absolutely cannot remove any <a href='#...'>...</a> tags, those are extremely important for references and are coming directly from the document, you MUST always preserve them.
 
 **Example:**
 
@@ -58,7 +59,7 @@ class LLMTextProcessor(BaseLLMSimpleBlockProcessor):
 ```
 
 Output:
-
+analysis: The inline math in the lines is not in LaTeX format and is not surrounded by <math>...</math> tags.
 ```json
 {
  "corrected_lines": [
@@ -163,4 +164,5 @@ def rewrite_block(self, response: dict, prompt_data: PromptData, document: Docum
             add_math_spans_to_line(corrected_text, text_line, page)
 
 class LLMTextSchema(BaseModel):
+    analysis: str
     corrected_lines: List[str]
\ No newline at end of file
diff --git a/marker/renderers/markdown.py b/marker/renderers/markdown.py
index 62463273..668377ed 100644
--- a/marker/renderers/markdown.py
+++ b/marker/renderers/markdown.py
@@ -81,11 +81,12 @@ def convert_p(self, el, text, convert_as_inline):
         return f"{text}\n\n" if text else ""  # default convert_p behavior
 
     def convert_math(self, el, text, convert_as_inline):
-        inline = el.has_attr('display') and el['display'] == 'inline'
-        if inline:
-            return " " + self.inline_math_delimiters[0] + text + self.inline_math_delimiters[1] + " "
-        else:
+        block = (el.has_attr('display') and el['display'] == 'block')
+        if block:
             return "\n" + self.block_math_delimiters[0] + text + self.block_math_delimiters[1] + "\n"
+        else:
+            return " " + self.inline_math_delimiters[0] + text + self.inline_math_delimiters[1] + " "
+
 
     def convert_table(self, el, text, convert_as_inline):
         total_rows = len(el.find_all('tr'))
diff --git a/marker/services/claude.py b/marker/services/claude.py
index 3dcd84b8..cb1055da 100644
--- a/marker/services/claude.py
+++ b/marker/services/claude.py
@@ -17,7 +17,7 @@ class ClaudeService(BaseService):
     claude_model_name: Annotated[
         str,
         "The name of the Google model to use for the service."
-    ] = "claude-3-5-haiku-20241022"
+    ] = "claude-3-7-sonnet-20250219"
     claude_api_key: Annotated[
         str,
         "The Claude API key to use for the service."
diff --git a/poetry.lock b/poetry.lock
index eb5298b1..9045cf29 100644
--- a/poetry.lock
+++ b/poetry.lock
@@ -431,6 +431,179 @@ files = [
     {file = "blinker-1.9.0.tar.gz", hash = "sha256:b4ce2265a7abece45e7cc896e98dbebe6cead56bcf805a3d23136d145f5445bf"},
 ]
 
+[[package]]
+name = "brotli"
+version = "1.1.0"
+description = "Python bindings for the Brotli compression library"
+optional = true
+python-versions = "*"
+files = [
+    {file = "Brotli-1.1.0-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:e1140c64812cb9b06c922e77f1c26a75ec5e3f0fb2bf92cc8c58720dec276752"},
+    {file = "Brotli-1.1.0-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:c8fd5270e906eef71d4a8d19b7c6a43760c6abcfcc10c9101d14eb2357418de9"},
+    {file = "Brotli-1.1.0-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:1ae56aca0402a0f9a3431cddda62ad71666ca9d4dc3a10a142b9dce2e3c0cda3"},
+    {file = "Brotli-1.1.0-cp310-cp310-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:43ce1b9935bfa1ede40028054d7f48b5469cd02733a365eec8a329ffd342915d"},
+    {file = "Brotli-1.1.0-cp310-cp310-manylinux_2_5_i686.manylinux1_i686.manylinux_2_12_i686.manylinux2010_i686.whl", hash = "sha256:7c4855522edb2e6ae7fdb58e07c3ba9111e7621a8956f481c68d5d979c93032e"},
+    {file = "Brotli-1.1.0-cp310-cp310-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_12_x86_64.manylinux2010_x86_64.whl", hash = "sha256:38025d9f30cf4634f8309c6874ef871b841eb3c347e90b0851f63d1ded5212da"},
+    {file = "Brotli-1.1.0-cp310-cp310-musllinux_1_1_aarch64.whl", hash = "sha256:e6a904cb26bfefc2f0a6f240bdf5233be78cd2488900a2f846f3c3ac8489ab80"},
+    {file = "Brotli-1.1.0-cp310-cp310-musllinux_1_1_i686.whl", hash = "sha256:a37b8f0391212d29b3a91a799c8e4a2855e0576911cdfb2515487e30e322253d"},
+    {file = "Brotli-1.1.0-cp310-cp310-musllinux_1_1_ppc64le.whl", hash = "sha256:e84799f09591700a4154154cab9787452925578841a94321d5ee8fb9a9a328f0"},
+    {file = "Brotli-1.1.0-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:f66b5337fa213f1da0d9000bc8dc0cb5b896b726eefd9c6046f699b169c41b9e"},
+    {file = "Brotli-1.1.0-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:5dab0844f2cf82be357a0eb11a9087f70c5430b2c241493fc122bb6f2bb0917c"},
+    {file = "Brotli-1.1.0-cp310-cp310-musllinux_1_2_i686.whl", hash = "sha256:e4fe605b917c70283db7dfe5ada75e04561479075761a0b3866c081d035b01c1"},
+    {file = "Brotli-1.1.0-cp310-cp310-musllinux_1_2_ppc64le.whl", hash = "sha256:1e9a65b5736232e7a7f91ff3d02277f11d339bf34099a56cdab6a8b3410a02b2"},
+    {file = "Brotli-1.1.0-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:58d4b711689366d4a03ac7957ab8c28890415e267f9b6589969e74b6e42225ec"},
+    {file = "Brotli-1.1.0-cp310-cp310-win32.whl", hash = "sha256:be36e3d172dc816333f33520154d708a2657ea63762ec16b62ece02ab5e4daf2"},
+    {file = "Brotli-1.1.0-cp310-cp310-win_amd64.whl", hash = "sha256:0c6244521dda65ea562d5a69b9a26120769b7a9fb3db2fe9545935ed6735b128"},
+    {file = "Brotli-1.1.0-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:a3daabb76a78f829cafc365531c972016e4aa8d5b4bf60660ad8ecee19df7ccc"},
+    {file = "Brotli-1.1.0-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:c8146669223164fc87a7e3de9f81e9423c67a79d6b3447994dfb9c95da16e2d6"},
+    {file = "Brotli-1.1.0-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:30924eb4c57903d5a7526b08ef4a584acc22ab1ffa085faceb521521d2de32dd"},
+    {file = "Brotli-1.1.0-cp311-cp311-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:ceb64bbc6eac5a140ca649003756940f8d6a7c444a68af170b3187623b43bebf"},
+    {file = "Brotli-1.1.0-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:a469274ad18dc0e4d316eefa616d1d0c2ff9da369af19fa6f3daa4f09671fd61"},
+    {file = "Brotli-1.1.0-cp311-cp311-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:524f35912131cc2cabb00edfd8d573b07f2d9f21fa824bd3fb19725a9cf06327"},
+    {file = "Brotli-1.1.0-cp311-cp311-musllinux_1_1_aarch64.whl", hash = "sha256:5b3cc074004d968722f51e550b41a27be656ec48f8afaeeb45ebf65b561481dd"},
+    {file = "Brotli-1.1.0-cp311-cp311-musllinux_1_1_i686.whl", hash = "sha256:19c116e796420b0cee3da1ccec3b764ed2952ccfcc298b55a10e5610ad7885f9"},
+    {file = "Brotli-1.1.0-cp311-cp311-musllinux_1_1_ppc64le.whl", hash = "sha256:510b5b1bfbe20e1a7b3baf5fed9e9451873559a976c1a78eebaa3b86c57b4265"},
+    {file = "Brotli-1.1.0-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:a1fd8a29719ccce974d523580987b7f8229aeace506952fa9ce1d53a033873c8"},
+    {file = "Brotli-1.1.0-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:c247dd99d39e0338a604f8c2b3bc7061d5c2e9e2ac7ba9cc1be5a69cb6cd832f"},
+    {file = "Brotli-1.1.0-cp311-cp311-musllinux_1_2_i686.whl", hash = "sha256:1b2c248cd517c222d89e74669a4adfa5577e06ab68771a529060cf5a156e9757"},
+    {file = "Brotli-1.1.0-cp311-cp311-musllinux_1_2_ppc64le.whl", hash = "sha256:2a24c50840d89ded6c9a8fdc7b6ed3692ed4e86f1c4a4a938e1e92def92933e0"},
+    {file = "Brotli-1.1.0-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:f31859074d57b4639318523d6ffdca586ace54271a73ad23ad021acd807eb14b"},
+    {file = "Brotli-1.1.0-cp311-cp311-win32.whl", hash = "sha256:39da8adedf6942d76dc3e46653e52df937a3c4d6d18fdc94a7c29d263b1f5b50"},
+    {file = "Brotli-1.1.0-cp311-cp311-win_amd64.whl", hash = "sha256:aac0411d20e345dc0920bdec5548e438e999ff68d77564d5e9463a7ca9d3e7b1"},
+    {file = "Brotli-1.1.0-cp312-cp312-macosx_10_13_universal2.whl", hash = "sha256:32d95b80260d79926f5fab3c41701dbb818fde1c9da590e77e571eefd14abe28"},
+    {file = "Brotli-1.1.0-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:b760c65308ff1e462f65d69c12e4ae085cff3b332d894637f6273a12a482d09f"},
+    {file = "Brotli-1.1.0-cp312-cp312-macosx_10_9_universal2.whl", hash = "sha256:316cc9b17edf613ac76b1f1f305d2a748f1b976b033b049a6ecdfd5612c70409"},
+    {file = "Brotli-1.1.0-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:caf9ee9a5775f3111642d33b86237b05808dafcd6268faa492250e9b78046eb2"},
+    {file = "Brotli-1.1.0-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:70051525001750221daa10907c77830bc889cb6d865cc0b813d9db7fefc21451"},
+    {file = "Brotli-1.1.0-cp312-cp312-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:7f4bf76817c14aa98cc6697ac02f3972cb8c3da93e9ef16b9c66573a68014f91"},
+    {file = "Brotli-1.1.0-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:d0c5516f0aed654134a2fc936325cc2e642f8a0e096d075209672eb321cff408"},
+    {file = "Brotli-1.1.0-cp312-cp312-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:6c3020404e0b5eefd7c9485ccf8393cfb75ec38ce75586e046573c9dc29967a0"},
+    {file = "Brotli-1.1.0-cp312-cp312-musllinux_1_1_aarch64.whl", hash = "sha256:4ed11165dd45ce798d99a136808a794a748d5dc38511303239d4e2363c0695dc"},
+    {file = "Brotli-1.1.0-cp312-cp312-musllinux_1_1_i686.whl", hash = "sha256:4093c631e96fdd49e0377a9c167bfd75b6d0bad2ace734c6eb20b348bc3ea180"},
+    {file = "Brotli-1.1.0-cp312-cp312-musllinux_1_1_ppc64le.whl", hash = "sha256:7e4c4629ddad63006efa0ef968c8e4751c5868ff0b1c5c40f76524e894c50248"},
+    {file = "Brotli-1.1.0-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:861bf317735688269936f755fa136a99d1ed526883859f86e41a5d43c61d8966"},
+    {file = "Brotli-1.1.0-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:87a3044c3a35055527ac75e419dfa9f4f3667a1e887ee80360589eb8c90aabb9"},
+    {file = "Brotli-1.1.0-cp312-cp312-musllinux_1_2_i686.whl", hash = "sha256:c5529b34c1c9d937168297f2c1fde7ebe9ebdd5e121297ff9c043bdb2ae3d6fb"},
+    {file = "Brotli-1.1.0-cp312-cp312-musllinux_1_2_ppc64le.whl", hash = "sha256:ca63e1890ede90b2e4454f9a65135a4d387a4585ff8282bb72964fab893f2111"},
+    {file = "Brotli-1.1.0-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:e79e6520141d792237c70bcd7a3b122d00f2613769ae0cb61c52e89fd3443839"},
+    {file = "Brotli-1.1.0-cp312-cp312-win32.whl", hash = "sha256:5f4d5ea15c9382135076d2fb28dde923352fe02951e66935a9efaac8f10e81b0"},
+    {file = "Brotli-1.1.0-cp312-cp312-win_amd64.whl", hash = "sha256:906bc3a79de8c4ae5b86d3d75a8b77e44404b0f4261714306e3ad248d8ab0951"},
+    {file = "Brotli-1.1.0-cp313-cp313-macosx_10_13_universal2.whl", hash = "sha256:8bf32b98b75c13ec7cf774164172683d6e7891088f6316e54425fde1efc276d5"},
+    {file = "Brotli-1.1.0-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:7bc37c4d6b87fb1017ea28c9508b36bbcb0c3d18b4260fcdf08b200c74a6aee8"},
+    {file = "Brotli-1.1.0-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:3c0ef38c7a7014ffac184db9e04debe495d317cc9c6fb10071f7fefd93100a4f"},
+    {file = "Brotli-1.1.0-cp313-cp313-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:91d7cc2a76b5567591d12c01f019dd7afce6ba8cba6571187e21e2fc418ae648"},
+    {file = "Brotli-1.1.0-cp313-cp313-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:a93dde851926f4f2678e704fadeb39e16c35d8baebd5252c9fd94ce8ce68c4a0"},
+    {file = "Brotli-1.1.0-cp313-cp313-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:f0db75f47be8b8abc8d9e31bc7aad0547ca26f24a54e6fd10231d623f183d089"},
+    {file = "Brotli-1.1.0-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:6967ced6730aed543b8673008b5a391c3b1076d834ca438bbd70635c73775368"},
+    {file = "Brotli-1.1.0-cp313-cp313-musllinux_1_2_i686.whl", hash = "sha256:7eedaa5d036d9336c95915035fb57422054014ebdeb6f3b42eac809928e40d0c"},
+    {file = "Brotli-1.1.0-cp313-cp313-musllinux_1_2_ppc64le.whl", hash = "sha256:d487f5432bf35b60ed625d7e1b448e2dc855422e87469e3f450aa5552b0eb284"},
+    {file = "Brotli-1.1.0-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:832436e59afb93e1836081a20f324cb185836c617659b07b129141a8426973c7"},
+    {file = "Brotli-1.1.0-cp313-cp313-win32.whl", hash = "sha256:43395e90523f9c23a3d5bdf004733246fba087f2948f87ab28015f12359ca6a0"},
+    {file = "Brotli-1.1.0-cp313-cp313-win_amd64.whl", hash = "sha256:9011560a466d2eb3f5a6e4929cf4a09be405c64154e12df0dd72713f6500e32b"},
+    {file = "Brotli-1.1.0-cp36-cp36m-macosx_10_9_x86_64.whl", hash = "sha256:a090ca607cbb6a34b0391776f0cb48062081f5f60ddcce5d11838e67a01928d1"},
+    {file = "Brotli-1.1.0-cp36-cp36m-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:2de9d02f5bda03d27ede52e8cfe7b865b066fa49258cbab568720aa5be80a47d"},
+    {file = "Brotli-1.1.0-cp36-cp36m-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:2333e30a5e00fe0fe55903c8832e08ee9c3b1382aacf4db26664a16528d51b4b"},
+    {file = "Brotli-1.1.0-cp36-cp36m-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:4d4a848d1837973bf0f4b5e54e3bec977d99be36a7895c61abb659301b02c112"},
+    {file = "Brotli-1.1.0-cp36-cp36m-manylinux_2_5_x86_64.manylinux1_x86_64.whl", hash = "sha256:fdc3ff3bfccdc6b9cc7c342c03aa2400683f0cb891d46e94b64a197910dc4064"},
+    {file = "Brotli-1.1.0-cp36-cp36m-musllinux_1_1_aarch64.whl", hash = "sha256:5eeb539606f18a0b232d4ba45adccde4125592f3f636a6182b4a8a436548b914"},
+    {file = "Brotli-1.1.0-cp36-cp36m-musllinux_1_1_i686.whl", hash = "sha256:fd5f17ff8f14003595ab414e45fce13d073e0762394f957182e69035c9f3d7c2"},
+    {file = "Brotli-1.1.0-cp36-cp36m-musllinux_1_1_ppc64le.whl", hash = "sha256:069a121ac97412d1fe506da790b3e69f52254b9df4eb665cd42460c837193354"},
+    {file = "Brotli-1.1.0-cp36-cp36m-musllinux_1_1_x86_64.whl", hash = "sha256:e93dfc1a1165e385cc8239fab7c036fb2cd8093728cbd85097b284d7b99249a2"},
+    {file = "Brotli-1.1.0-cp36-cp36m-musllinux_1_2_aarch64.whl", hash = "sha256:aea440a510e14e818e67bfc4027880e2fb500c2ccb20ab21c7a7c8b5b4703d75"},
+    {file = "Brotli-1.1.0-cp36-cp36m-musllinux_1_2_i686.whl", hash = "sha256:6974f52a02321b36847cd19d1b8e381bf39939c21efd6ee2fc13a28b0d99348c"},
+    {file = "Brotli-1.1.0-cp36-cp36m-musllinux_1_2_ppc64le.whl", hash = "sha256:a7e53012d2853a07a4a79c00643832161a910674a893d296c9f1259859a289d2"},
+    {file = "Brotli-1.1.0-cp36-cp36m-musllinux_1_2_x86_64.whl", hash = "sha256:d7702622a8b40c49bffb46e1e3ba2e81268d5c04a34f460978c6b5517a34dd52"},
+    {file = "Brotli-1.1.0-cp36-cp36m-win32.whl", hash = "sha256:a599669fd7c47233438a56936988a2478685e74854088ef5293802123b5b2460"},
+    {file = "Brotli-1.1.0-cp36-cp36m-win_amd64.whl", hash = "sha256:d143fd47fad1db3d7c27a1b1d66162e855b5d50a89666af46e1679c496e8e579"},
+    {file = "Brotli-1.1.0-cp37-cp37m-macosx_10_9_x86_64.whl", hash = "sha256:11d00ed0a83fa22d29bc6b64ef636c4552ebafcef57154b4ddd132f5638fbd1c"},
+    {file = "Brotli-1.1.0-cp37-cp37m-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:f733d788519c7e3e71f0855c96618720f5d3d60c3cb829d8bbb722dddce37985"},
+    {file = "Brotli-1.1.0-cp37-cp37m-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:929811df5462e182b13920da56c6e0284af407d1de637d8e536c5cd00a7daf60"},
+    {file = "Brotli-1.1.0-cp37-cp37m-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:0b63b949ff929fbc2d6d3ce0e924c9b93c9785d877a21a1b678877ffbbc4423a"},
+    {file = "Brotli-1.1.0-cp37-cp37m-manylinux_2_5_x86_64.manylinux1_x86_64.whl", hash = "sha256:d192f0f30804e55db0d0e0a35d83a9fead0e9a359a9ed0285dbacea60cc10a84"},
+    {file = "Brotli-1.1.0-cp37-cp37m-musllinux_1_1_aarch64.whl", hash = "sha256:f296c40e23065d0d6650c4aefe7470d2a25fffda489bcc3eb66083f3ac9f6643"},
+    {file = "Brotli-1.1.0-cp37-cp37m-musllinux_1_1_i686.whl", hash = "sha256:919e32f147ae93a09fe064d77d5ebf4e35502a8df75c29fb05788528e330fe74"},
+    {file = "Brotli-1.1.0-cp37-cp37m-musllinux_1_1_ppc64le.whl", hash = "sha256:23032ae55523cc7bccb4f6a0bf368cd25ad9bcdcc1990b64a647e7bbcce9cb5b"},
+    {file = "Brotli-1.1.0-cp37-cp37m-musllinux_1_1_x86_64.whl", hash = "sha256:224e57f6eac61cc449f498cc5f0e1725ba2071a3d4f48d5d9dffba42db196438"},
+    {file = "Brotli-1.1.0-cp37-cp37m-musllinux_1_2_aarch64.whl", hash = "sha256:cb1dac1770878ade83f2ccdf7d25e494f05c9165f5246b46a621cc849341dc01"},
+    {file = "Brotli-1.1.0-cp37-cp37m-musllinux_1_2_i686.whl", hash = "sha256:3ee8a80d67a4334482d9712b8e83ca6b1d9bc7e351931252ebef5d8f7335a547"},
+    {file = "Brotli-1.1.0-cp37-cp37m-musllinux_1_2_ppc64le.whl", hash = "sha256:5e55da2c8724191e5b557f8e18943b1b4839b8efc3ef60d65985bcf6f587dd38"},
+    {file = "Brotli-1.1.0-cp37-cp37m-musllinux_1_2_x86_64.whl", hash = "sha256:d342778ef319e1026af243ed0a07c97acf3bad33b9f29e7ae6a1f68fd083e90c"},
+    {file = "Brotli-1.1.0-cp37-cp37m-win32.whl", hash = "sha256:587ca6d3cef6e4e868102672d3bd9dc9698c309ba56d41c2b9c85bbb903cdb95"},
+    {file = "Brotli-1.1.0-cp37-cp37m-win_amd64.whl", hash = "sha256:2954c1c23f81c2eaf0b0717d9380bd348578a94161a65b3a2afc62c86467dd68"},
+    {file = "Brotli-1.1.0-cp38-cp38-macosx_10_9_universal2.whl", hash = "sha256:efa8b278894b14d6da122a72fefcebc28445f2d3f880ac59d46c90f4c13be9a3"},
+    {file = "Brotli-1.1.0-cp38-cp38-macosx_10_9_x86_64.whl", hash = "sha256:03d20af184290887bdea3f0f78c4f737d126c74dc2f3ccadf07e54ceca3bf208"},
+    {file = "Brotli-1.1.0-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:6172447e1b368dcbc458925e5ddaf9113477b0ed542df258d84fa28fc45ceea7"},
+    {file = "Brotli-1.1.0-cp38-cp38-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:a743e5a28af5f70f9c080380a5f908d4d21d40e8f0e0c8901604d15cfa9ba751"},
+    {file = "Brotli-1.1.0-cp38-cp38-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:0541e747cce78e24ea12d69176f6a7ddb690e62c425e01d31cc065e69ce55b48"},
+    {file = "Brotli-1.1.0-cp38-cp38-manylinux_2_5_x86_64.manylinux1_x86_64.whl", hash = "sha256:cdbc1fc1bc0bff1cef838eafe581b55bfbffaed4ed0318b724d0b71d4d377619"},
+    {file = "Brotli-1.1.0-cp38-cp38-musllinux_1_1_aarch64.whl", hash = "sha256:890b5a14ce214389b2cc36ce82f3093f96f4cc730c1cffdbefff77a7c71f2a97"},
+    {file = "Brotli-1.1.0-cp38-cp38-musllinux_1_1_i686.whl", hash = "sha256:1ab4fbee0b2d9098c74f3057b2bc055a8bd92ccf02f65944a241b4349229185a"},
+    {file = "Brotli-1.1.0-cp38-cp38-musllinux_1_1_ppc64le.whl", hash = "sha256:141bd4d93984070e097521ed07e2575b46f817d08f9fa42b16b9b5f27b5ac088"},
+    {file = "Brotli-1.1.0-cp38-cp38-musllinux_1_1_x86_64.whl", hash = "sha256:fce1473f3ccc4187f75b4690cfc922628aed4d3dd013d047f95a9b3919a86596"},
+    {file = "Brotli-1.1.0-cp38-cp38-musllinux_1_2_aarch64.whl", hash = "sha256:d2b35ca2c7f81d173d2fadc2f4f31e88cc5f7a39ae5b6db5513cf3383b0e0ec7"},
+    {file = "Brotli-1.1.0-cp38-cp38-musllinux_1_2_i686.whl", hash = "sha256:af6fa6817889314555aede9a919612b23739395ce767fe7fcbea9a80bf140fe5"},
+    {file = "Brotli-1.1.0-cp38-cp38-musllinux_1_2_ppc64le.whl", hash = "sha256:2feb1d960f760a575dbc5ab3b1c00504b24caaf6986e2dc2b01c09c87866a943"},
+    {file = "Brotli-1.1.0-cp38-cp38-musllinux_1_2_x86_64.whl", hash = "sha256:4410f84b33374409552ac9b6903507cdb31cd30d2501fc5ca13d18f73548444a"},
+    {file = "Brotli-1.1.0-cp38-cp38-win32.whl", hash = "sha256:db85ecf4e609a48f4b29055f1e144231b90edc90af7481aa731ba2d059226b1b"},
+    {file = "Brotli-1.1.0-cp38-cp38-win_amd64.whl", hash = "sha256:3d7954194c36e304e1523f55d7042c59dc53ec20dd4e9ea9d151f1b62b4415c0"},
+    {file = "Brotli-1.1.0-cp39-cp39-macosx_10_9_universal2.whl", hash = "sha256:5fb2ce4b8045c78ebbc7b8f3c15062e435d47e7393cc57c25115cfd49883747a"},
+    {file = "Brotli-1.1.0-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:7905193081db9bfa73b1219140b3d315831cbff0d8941f22da695832f0dd188f"},
+    {file = "Brotli-1.1.0-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:a77def80806c421b4b0af06f45d65a136e7ac0bdca3c09d9e2ea4e515367c7e9"},
+    {file = "Brotli-1.1.0-cp39-cp39-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:8dadd1314583ec0bf2d1379f7008ad627cd6336625d6679cf2f8e67081b83acf"},
+    {file = "Brotli-1.1.0-cp39-cp39-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:901032ff242d479a0efa956d853d16875d42157f98951c0230f69e69f9c09bac"},
+    {file = "Brotli-1.1.0-cp39-cp39-manylinux_2_5_x86_64.manylinux1_x86_64.whl", hash = "sha256:22fc2a8549ffe699bfba2256ab2ed0421a7b8fadff114a3d201794e45a9ff578"},
+    {file = "Brotli-1.1.0-cp39-cp39-musllinux_1_1_aarch64.whl", hash = "sha256:ae15b066e5ad21366600ebec29a7ccbc86812ed267e4b28e860b8ca16a2bc474"},
+    {file = "Brotli-1.1.0-cp39-cp39-musllinux_1_1_i686.whl", hash = "sha256:949f3b7c29912693cee0afcf09acd6ebc04c57af949d9bf77d6101ebb61e388c"},
+    {file = "Brotli-1.1.0-cp39-cp39-musllinux_1_1_ppc64le.whl", hash = "sha256:89f4988c7203739d48c6f806f1e87a1d96e0806d44f0fba61dba81392c9e474d"},
+    {file = "Brotli-1.1.0-cp39-cp39-musllinux_1_1_x86_64.whl", hash = "sha256:de6551e370ef19f8de1807d0a9aa2cdfdce2e85ce88b122fe9f6b2b076837e59"},
+    {file = "Brotli-1.1.0-cp39-cp39-musllinux_1_2_aarch64.whl", hash = "sha256:0737ddb3068957cf1b054899b0883830bb1fec522ec76b1098f9b6e0f02d9419"},
+    {file = "Brotli-1.1.0-cp39-cp39-musllinux_1_2_i686.whl", hash = "sha256:4f3607b129417e111e30637af1b56f24f7a49e64763253bbc275c75fa887d4b2"},
+    {file = "Brotli-1.1.0-cp39-cp39-musllinux_1_2_ppc64le.whl", hash = "sha256:6c6e0c425f22c1c719c42670d561ad682f7bfeeef918edea971a79ac5252437f"},
+    {file = "Brotli-1.1.0-cp39-cp39-musllinux_1_2_x86_64.whl", hash = "sha256:494994f807ba0b92092a163a0a283961369a65f6cbe01e8891132b7a320e61eb"},
+    {file = "Brotli-1.1.0-cp39-cp39-win32.whl", hash = "sha256:f0d8a7a6b5983c2496e364b969f0e526647a06b075d034f3297dc66f3b360c64"},
+    {file = "Brotli-1.1.0-cp39-cp39-win_amd64.whl", hash = "sha256:cdad5b9014d83ca68c25d2e9444e28e967ef16e80f6b436918c700c117a85467"},
+    {file = "Brotli-1.1.0.tar.gz", hash = "sha256:81de08ac11bcb85841e440c13611c00b67d3bf82698314928d0b676362546724"},
+]
+
+[[package]]
+name = "brotlicffi"
+version = "1.1.0.0"
+description = "Python CFFI bindings to the Brotli library"
+optional = true
+python-versions = ">=3.7"
+files = [
+    {file = "brotlicffi-1.1.0.0-cp37-abi3-macosx_10_9_x86_64.whl", hash = "sha256:9b7ae6bd1a3f0df532b6d67ff674099a96d22bc0948955cb338488c31bfb8851"},
+    {file = "brotlicffi-1.1.0.0-cp37-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:19ffc919fa4fc6ace69286e0a23b3789b4219058313cf9b45625016bf7ff996b"},
+    {file = "brotlicffi-1.1.0.0-cp37-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:9feb210d932ffe7798ee62e6145d3a757eb6233aa9a4e7db78dd3690d7755814"},
+    {file = "brotlicffi-1.1.0.0-cp37-abi3-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:84763dbdef5dd5c24b75597a77e1b30c66604725707565188ba54bab4f114820"},
+    {file = "brotlicffi-1.1.0.0-cp37-abi3-win32.whl", hash = "sha256:1b12b50e07c3911e1efa3a8971543e7648100713d4e0971b13631cce22c587eb"},
+    {file = "brotlicffi-1.1.0.0-cp37-abi3-win_amd64.whl", hash = "sha256:994a4f0681bb6c6c3b0925530a1926b7a189d878e6e5e38fae8efa47c5d9c613"},
+    {file = "brotlicffi-1.1.0.0-pp310-pypy310_pp73-macosx_10_9_x86_64.whl", hash = "sha256:2e4aeb0bd2540cb91b069dbdd54d458da8c4334ceaf2d25df2f4af576d6766ca"},
+    {file = "brotlicffi-1.1.0.0-pp310-pypy310_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:4b7b0033b0d37bb33009fb2fef73310e432e76f688af76c156b3594389d81391"},
+    {file = "brotlicffi-1.1.0.0-pp310-pypy310_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:54a07bb2374a1eba8ebb52b6fafffa2afd3c4df85ddd38fcc0511f2bb387c2a8"},
+    {file = "brotlicffi-1.1.0.0-pp310-pypy310_pp73-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:7901a7dc4b88f1c1475de59ae9be59799db1007b7d059817948d8e4f12e24e35"},
+    {file = "brotlicffi-1.1.0.0-pp310-pypy310_pp73-win_amd64.whl", hash = "sha256:ce01c7316aebc7fce59da734286148b1d1b9455f89cf2c8a4dfce7d41db55c2d"},
+    {file = "brotlicffi-1.1.0.0-pp37-pypy37_pp73-macosx_10_9_x86_64.whl", hash = "sha256:246f1d1a90279bb6069de3de8d75a8856e073b8ff0b09dcca18ccc14cec85979"},
+    {file = "brotlicffi-1.1.0.0-pp37-pypy37_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:cc4bc5d82bc56ebd8b514fb8350cfac4627d6b0743382e46d033976a5f80fab6"},
+    {file = "brotlicffi-1.1.0.0-pp37-pypy37_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:37c26ecb14386a44b118ce36e546ce307f4810bc9598a6e6cb4f7fca725ae7e6"},
+    {file = "brotlicffi-1.1.0.0-pp37-pypy37_pp73-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:ca72968ae4eaf6470498d5c2887073f7efe3b1e7d7ec8be11a06a79cc810e990"},
+    {file = "brotlicffi-1.1.0.0-pp37-pypy37_pp73-win_amd64.whl", hash = "sha256:add0de5b9ad9e9aa293c3aa4e9deb2b61e99ad6c1634e01d01d98c03e6a354cc"},
+    {file = "brotlicffi-1.1.0.0-pp38-pypy38_pp73-macosx_10_9_x86_64.whl", hash = "sha256:9b6068e0f3769992d6b622a1cd2e7835eae3cf8d9da123d7f51ca9c1e9c333e5"},
+    {file = "brotlicffi-1.1.0.0-pp38-pypy38_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:8557a8559509b61e65083f8782329188a250102372576093c88930c875a69838"},
+    {file = "brotlicffi-1.1.0.0-pp38-pypy38_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:2a7ae37e5d79c5bdfb5b4b99f2715a6035e6c5bf538c3746abc8e26694f92f33"},
+    {file = "brotlicffi-1.1.0.0-pp38-pypy38_pp73-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:391151ec86bb1c683835980f4816272a87eaddc46bb91cbf44f62228b84d8cca"},
+    {file = "brotlicffi-1.1.0.0-pp38-pypy38_pp73-win_amd64.whl", hash = "sha256:2f3711be9290f0453de8eed5275d93d286abe26b08ab4a35d7452caa1fef532f"},
+    {file = "brotlicffi-1.1.0.0-pp39-pypy39_pp73-macosx_10_9_x86_64.whl", hash = "sha256:1a807d760763e398bbf2c6394ae9da5815901aa93ee0a37bca5efe78d4ee3171"},
+    {file = "brotlicffi-1.1.0.0-pp39-pypy39_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:fa8ca0623b26c94fccc3a1fdd895be1743b838f3917300506d04aa3346fd2a14"},
+    {file = "brotlicffi-1.1.0.0-pp39-pypy39_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:3de0cf28a53a3238b252aca9fed1593e9d36c1d116748013339f0949bfc84112"},
+    {file = "brotlicffi-1.1.0.0-pp39-pypy39_pp73-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:6be5ec0e88a4925c91f3dea2bb0013b3a2accda6f77238f76a34a1ea532a1cb0"},
+    {file = "brotlicffi-1.1.0.0-pp39-pypy39_pp73-win_amd64.whl", hash = "sha256:d9eb71bb1085d996244439154387266fd23d6ad37161f6f52f1cd41dd95a3808"},
+    {file = "brotlicffi-1.1.0.0.tar.gz", hash = "sha256:b77827a689905143f87915310b93b273ab17888fd43ef350d4832c4a71083c13"},
+]
+
+[package.dependencies]
+cffi = ">=1.0.0"
+
 [[package]]
 name = "cachetools"
 version = "5.5.2"
@@ -651,7 +824,7 @@ colorama = {version = "*", markers = "platform_system == \"Windows\""}
 name = "cobble"
 version = "0.1.4"
 description = "Create data objects"
-optional = false
+optional = true
 python-versions = ">=3.5"
 files = [
     {file = "cobble-0.1.4-py3-none-any.whl", hash = "sha256:36c91b1655e599fd428e2b95fdd5f0da1ca2e9f1abb0bc871dec21a0e78a2b44"},
@@ -686,6 +859,25 @@ traitlets = ">=4"
 [package.extras]
 test = ["pytest"]
 
+[[package]]
+name = "cssselect2"
+version = "0.7.0"
+description = "CSS selectors for Python ElementTree"
+optional = true
+python-versions = ">=3.7"
+files = [
+    {file = "cssselect2-0.7.0-py3-none-any.whl", hash = "sha256:fd23a65bfd444595913f02fc71f6b286c29261e354c41d722ca7a261a49b5969"},
+    {file = "cssselect2-0.7.0.tar.gz", hash = "sha256:1ccd984dab89fc68955043aca4e1b03e0cf29cad9880f6e28e3ba7a74b14aa5a"},
+]
+
+[package.dependencies]
+tinycss2 = "*"
+webencodings = "*"
+
+[package.extras]
+doc = ["sphinx", "sphinx_rtd_theme"]
+test = ["flake8", "isort", "pytest"]
+
 [[package]]
 name = "datasets"
 version = "2.21.0"
@@ -827,7 +1019,7 @@ files = [
 name = "ebooklib"
 version = "0.18"
 description = "Ebook library which can handle EPUB2/EPUB3 and Kindle format"
-optional = false
+optional = true
 python-versions = "*"
 files = [
     {file = "EbookLib-0.18.tar.gz", hash = "sha256:38562643a7bc94d9bf56e9930b4927e4e93b5d1d0917f697a6454db5a1c1a533"},
@@ -841,7 +1033,7 @@ six = "*"
 name = "et-xmlfile"
 version = "2.0.0"
 description = "An implementation of lxml.xmlfile for the standard library"
-optional = false
+optional = true
 python-versions = ">=3.8"
 files = [
     {file = "et_xmlfile-2.0.0-py3-none-any.whl", hash = "sha256:7a91720bc756843502c3b7504c77b8fe44217c85c537d85037f0f536151b2caa"},
@@ -878,18 +1070,18 @@ tests = ["asttokens (>=2.1.0)", "coverage", "coverage-enable-subprocess", "ipyth
 
 [[package]]
 name = "fastapi"
-version = "0.115.8"
+version = "0.115.10"
 description = "FastAPI framework, high performance, easy to learn, fast to code, ready for production"
 optional = false
 python-versions = ">=3.8"
 files = [
-    {file = "fastapi-0.115.8-py3-none-any.whl", hash = "sha256:753a96dd7e036b34eeef8babdfcfe3f28ff79648f86551eb36bfc1b0bf4a8cbf"},
-    {file = "fastapi-0.115.8.tar.gz", hash = "sha256:0ce9111231720190473e222cdf0f07f7206ad7e53ea02beb1d2dc36e2f0741e9"},
+    {file = "fastapi-0.115.10-py3-none-any.whl", hash = "sha256:47346c5437e933e68909a835cf63890a9bd52fb6091b2499b996c08a01ca43a5"},
+    {file = "fastapi-0.115.10.tar.gz", hash = "sha256:920cdc95c1c6ca073656deae80ad254512d131031c2d7759c87ae469572911ee"},
 ]
 
 [package.dependencies]
 pydantic = ">=1.7.4,<1.8 || >1.8,<1.8.1 || >1.8.1,<2.0.0 || >2.0.0,<2.0.1 || >2.0.1,<2.1.0 || >2.1.0,<3.0.0"
-starlette = ">=0.40.0,<0.46.0"
+starlette = ">=0.40.0,<0.47.0"
 typing-extensions = ">=4.8.0"
 
 [package.extras]
@@ -937,6 +1129,84 @@ files = [
     {file = "filetype-1.2.0.tar.gz", hash = "sha256:66b56cd6474bf41d8c54660347d37afcc3f7d1970648de365c102ef77548aadb"},
 ]
 
+[[package]]
+name = "fonttools"
+version = "4.56.0"
+description = "Tools to manipulate font files"
+optional = true
+python-versions = ">=3.8"
+files = [
+    {file = "fonttools-4.56.0-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:331954d002dbf5e704c7f3756028e21db07097c19722569983ba4d74df014000"},
+    {file = "fonttools-4.56.0-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:8d1613abd5af2f93c05867b3a3759a56e8bf97eb79b1da76b2bc10892f96ff16"},
+    {file = "fonttools-4.56.0-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:705837eae384fe21cee5e5746fd4f4b2f06f87544fa60f60740007e0aa600311"},
+    {file = "fonttools-4.56.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:bc871904a53a9d4d908673c6faa15689874af1c7c5ac403a8e12d967ebd0c0dc"},
+    {file = "fonttools-4.56.0-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:38b947de71748bab150259ee05a775e8a0635891568e9fdb3cdd7d0e0004e62f"},
+    {file = "fonttools-4.56.0-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:86b2a1013ef7a64d2e94606632683f07712045ed86d937c11ef4dde97319c086"},
+    {file = "fonttools-4.56.0-cp310-cp310-win32.whl", hash = "sha256:133bedb9a5c6376ad43e6518b7e2cd2f866a05b1998f14842631d5feb36b5786"},
+    {file = "fonttools-4.56.0-cp310-cp310-win_amd64.whl", hash = "sha256:17f39313b649037f6c800209984a11fc256a6137cbe5487091c6c7187cae4685"},
+    {file = "fonttools-4.56.0-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:7ef04bc7827adb7532be3d14462390dd71287644516af3f1e67f1e6ff9c6d6df"},
+    {file = "fonttools-4.56.0-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:ffda9b8cd9cb8b301cae2602ec62375b59e2e2108a117746f12215145e3f786c"},
+    {file = "fonttools-4.56.0-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:e2e993e8db36306cc3f1734edc8ea67906c55f98683d6fd34c3fc5593fdbba4c"},
+    {file = "fonttools-4.56.0-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:003548eadd674175510773f73fb2060bb46adb77c94854af3e0cc5bc70260049"},
+    {file = "fonttools-4.56.0-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:bd9825822e7bb243f285013e653f6741954d8147427aaa0324a862cdbf4cbf62"},
+    {file = "fonttools-4.56.0-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:b23d30a2c0b992fb1c4f8ac9bfde44b5586d23457759b6cf9a787f1a35179ee0"},
+    {file = "fonttools-4.56.0-cp311-cp311-win32.whl", hash = "sha256:47b5e4680002ae1756d3ae3b6114e20aaee6cc5c69d1e5911f5ffffd3ee46c6b"},
+    {file = "fonttools-4.56.0-cp311-cp311-win_amd64.whl", hash = "sha256:14a3e3e6b211660db54ca1ef7006401e4a694e53ffd4553ab9bc87ead01d0f05"},
+    {file = "fonttools-4.56.0-cp312-cp312-macosx_10_13_universal2.whl", hash = "sha256:d6f195c14c01bd057bc9b4f70756b510e009c83c5ea67b25ced3e2c38e6ee6e9"},
+    {file = "fonttools-4.56.0-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:fa760e5fe8b50cbc2d71884a1eff2ed2b95a005f02dda2fa431560db0ddd927f"},
+    {file = "fonttools-4.56.0-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:d54a45d30251f1d729e69e5b675f9a08b7da413391a1227781e2a297fa37f6d2"},
+    {file = "fonttools-4.56.0-cp312-cp312-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:661a8995d11e6e4914a44ca7d52d1286e2d9b154f685a4d1f69add8418961563"},
+    {file = "fonttools-4.56.0-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:9d94449ad0a5f2a8bf5d2f8d71d65088aee48adbe45f3c5f8e00e3ad861ed81a"},
+    {file = "fonttools-4.56.0-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:f59746f7953f69cc3290ce2f971ab01056e55ddd0fb8b792c31a8acd7fee2d28"},
+    {file = "fonttools-4.56.0-cp312-cp312-win32.whl", hash = "sha256:bce60f9a977c9d3d51de475af3f3581d9b36952e1f8fc19a1f2254f1dda7ce9c"},
+    {file = "fonttools-4.56.0-cp312-cp312-win_amd64.whl", hash = "sha256:300c310bb725b2bdb4f5fc7e148e190bd69f01925c7ab437b9c0ca3e1c7cd9ba"},
+    {file = "fonttools-4.56.0-cp313-cp313-macosx_10_13_universal2.whl", hash = "sha256:f20e2c0dfab82983a90f3d00703ac0960412036153e5023eed2b4641d7d5e692"},
+    {file = "fonttools-4.56.0-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:f36a0868f47b7566237640c026c65a86d09a3d9ca5df1cd039e30a1da73098a0"},
+    {file = "fonttools-4.56.0-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:62b4c6802fa28e14dba010e75190e0e6228513573f1eeae57b11aa1a39b7e5b1"},
+    {file = "fonttools-4.56.0-cp313-cp313-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:a05d1f07eb0a7d755fbe01fee1fd255c3a4d3730130cf1bfefb682d18fd2fcea"},
+    {file = "fonttools-4.56.0-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:0073b62c3438cf0058488c002ea90489e8801d3a7af5ce5f7c05c105bee815c3"},
+    {file = "fonttools-4.56.0-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:e2cad98c94833465bcf28f51c248aaf07ca022efc6a3eba750ad9c1e0256d278"},
+    {file = "fonttools-4.56.0-cp313-cp313-win32.whl", hash = "sha256:d0cb73ccf7f6d7ca8d0bc7ea8ac0a5b84969a41c56ac3ac3422a24df2680546f"},
+    {file = "fonttools-4.56.0-cp313-cp313-win_amd64.whl", hash = "sha256:62cc1253827d1e500fde9dbe981219fea4eb000fd63402283472d38e7d8aa1c6"},
+    {file = "fonttools-4.56.0-cp38-cp38-macosx_10_9_universal2.whl", hash = "sha256:3fd3fccb7b9adaaecfa79ad51b759f2123e1aba97f857936ce044d4f029abd71"},
+    {file = "fonttools-4.56.0-cp38-cp38-macosx_10_9_x86_64.whl", hash = "sha256:193b86e9f769320bc98ffdb42accafb5d0c8c49bd62884f1c0702bc598b3f0a2"},
+    {file = "fonttools-4.56.0-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:6e81c1cc80c1d8bf071356cc3e0e25071fbba1c75afc48d41b26048980b3c771"},
+    {file = "fonttools-4.56.0-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:e9270505a19361e81eecdbc2c251ad1e1a9a9c2ad75fa022ccdee533f55535dc"},
+    {file = "fonttools-4.56.0-cp38-cp38-musllinux_1_2_aarch64.whl", hash = "sha256:53f5e9767978a4daf46f28e09dbeb7d010319924ae622f7b56174b777258e5ba"},
+    {file = "fonttools-4.56.0-cp38-cp38-musllinux_1_2_x86_64.whl", hash = "sha256:9da650cb29bc098b8cfd15ef09009c914b35c7986c8fa9f08b51108b7bc393b4"},
+    {file = "fonttools-4.56.0-cp38-cp38-win32.whl", hash = "sha256:965d0209e6dbdb9416100123b6709cb13f5232e2d52d17ed37f9df0cc31e2b35"},
+    {file = "fonttools-4.56.0-cp38-cp38-win_amd64.whl", hash = "sha256:654ac4583e2d7c62aebc6fc6a4c6736f078f50300e18aa105d87ce8925cfac31"},
+    {file = "fonttools-4.56.0-cp39-cp39-macosx_10_9_universal2.whl", hash = "sha256:ca7962e8e5fc047cc4e59389959843aafbf7445b6c08c20d883e60ced46370a5"},
+    {file = "fonttools-4.56.0-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:a1af375734018951c31c0737d04a9d5fd0a353a0253db5fbed2ccd44eac62d8c"},
+    {file = "fonttools-4.56.0-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:442ad4122468d0e47d83bc59d0e91b474593a8c813839e1872e47c7a0cb53b10"},
+    {file = "fonttools-4.56.0-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:3cf4f8d2a30b454ac682e12c61831dcb174950c406011418e739de592bbf8f76"},
+    {file = "fonttools-4.56.0-cp39-cp39-musllinux_1_2_aarch64.whl", hash = "sha256:96a4271f63a615bcb902b9f56de00ea225d6896052c49f20d0c91e9f43529a29"},
+    {file = "fonttools-4.56.0-cp39-cp39-musllinux_1_2_x86_64.whl", hash = "sha256:6c1d38642ca2dddc7ae992ef5d026e5061a84f10ff2b906be5680ab089f55bb8"},
+    {file = "fonttools-4.56.0-cp39-cp39-win32.whl", hash = "sha256:2d351275f73ebdd81dd5b09a8b8dac7a30f29a279d41e1c1192aedf1b6dced40"},
+    {file = "fonttools-4.56.0-cp39-cp39-win_amd64.whl", hash = "sha256:d6ca96d1b61a707ba01a43318c9c40aaf11a5a568d1e61146fafa6ab20890793"},
+    {file = "fonttools-4.56.0-py3-none-any.whl", hash = "sha256:1088182f68c303b50ca4dc0c82d42083d176cba37af1937e1a976a31149d4d14"},
+    {file = "fonttools-4.56.0.tar.gz", hash = "sha256:a114d1567e1a1586b7e9e7fc2ff686ca542a82769a296cef131e4c4af51e58f4"},
+]
+
+[package.dependencies]
+brotli = {version = ">=1.0.1", optional = true, markers = "platform_python_implementation == \"CPython\" and extra == \"woff\""}
+brotlicffi = {version = ">=0.8.0", optional = true, markers = "platform_python_implementation != \"CPython\" and extra == \"woff\""}
+zopfli = {version = ">=0.1.4", optional = true, markers = "extra == \"woff\""}
+
+[package.extras]
+all = ["brotli (>=1.0.1)", "brotlicffi (>=0.8.0)", "fs (>=2.2.0,<3)", "lxml (>=4.0)", "lz4 (>=1.7.4.2)", "matplotlib", "munkres", "pycairo", "scipy", "skia-pathops (>=0.5.0)", "sympy", "uharfbuzz (>=0.23.0)", "unicodedata2 (>=15.1.0)", "xattr", "zopfli (>=0.1.4)"]
+graphite = ["lz4 (>=1.7.4.2)"]
+interpolatable = ["munkres", "pycairo", "scipy"]
+lxml = ["lxml (>=4.0)"]
+pathops = ["skia-pathops (>=0.5.0)"]
+plot = ["matplotlib"]
+repacker = ["uharfbuzz (>=0.23.0)"]
+symfont = ["sympy"]
+type1 = ["xattr"]
+ufo = ["fs (>=2.2.0,<3)"]
+unicode = ["unicodedata2 (>=15.1.0)"]
+woff = ["brotli (>=1.0.1)", "brotlicffi (>=0.8.0)", "zopfli (>=0.1.4)"]
+
 [[package]]
 name = "fqdn"
 version = "1.5.1"
@@ -1416,13 +1686,13 @@ test = ["flaky", "ipyparallel", "pre-commit", "pytest (>=7.0)", "pytest-asyncio
 
 [[package]]
 name = "ipython"
-version = "8.32.0"
+version = "8.33.0"
 description = "IPython: Productive Interactive Computing"
 optional = false
 python-versions = ">=3.10"
 files = [
-    {file = "ipython-8.32.0-py3-none-any.whl", hash = "sha256:cae85b0c61eff1fc48b0a8002de5958b6528fa9c8defb1894da63f42613708aa"},
-    {file = "ipython-8.32.0.tar.gz", hash = "sha256:be2c91895b0b9ea7ba49d33b23e2040c352b33eb6a519cca7ce6e0c743444251"},
+    {file = "ipython-8.33.0-py3-none-any.whl", hash = "sha256:aa5b301dfe1eaf0167ff3238a6825f810a029c9dad9d3f1597f30bd5ff65cc44"},
+    {file = "ipython-8.33.0.tar.gz", hash = "sha256:4c3e36a6dfa9e8e3702bd46f3df668624c975a22ff340e96ea7277afbd76217d"},
 ]
 
 [package.dependencies]
@@ -2117,7 +2387,7 @@ source = ["Cython (>=3.0.11)"]
 name = "mammoth"
 version = "1.9.0"
 description = "Convert Word documents from docx to simple and clean HTML and Markdown"
-optional = false
+optional = true
 python-versions = ">=3.7"
 files = [
     {file = "mammoth-1.9.0-py2.py3-none-any.whl", hash = "sha256:0eea277316586f0ca65d86834aec4de5a0572c83ec54b4991f9bb520a891150f"},
@@ -2863,17 +3133,17 @@ files = [
 
 [package.dependencies]
 numpy = [
+    {version = ">=1.26.0", markers = "python_version >= \"3.12\""},
+    {version = ">=1.23.5", markers = "python_version >= \"3.11\" and python_version < \"3.12\""},
     {version = ">=1.21.4", markers = "python_version >= \"3.10\" and platform_system == \"Darwin\" and python_version < \"3.11\""},
     {version = ">=1.21.2", markers = "platform_system != \"Darwin\" and python_version >= \"3.10\" and python_version < \"3.11\""},
-    {version = ">=1.23.5", markers = "python_version >= \"3.11\" and python_version < \"3.12\""},
-    {version = ">=1.26.0", markers = "python_version >= \"3.12\""},
 ]
 
 [[package]]
 name = "openpyxl"
 version = "3.1.5"
 description = "A Python library to read/write Excel 2010 xlsx/xlsm files"
-optional = false
+optional = true
 python-versions = ">=3.8"
 files = [
     {file = "openpyxl-3.1.5-py2.py3-none-any.whl", hash = "sha256:5282c12b107bffeef825f4617dc029afaf41d0ea60823bbb665ef3079dc79de2"},
@@ -2958,9 +3228,9 @@ files = [
 
 [package.dependencies]
 numpy = [
-    {version = ">=1.22.4", markers = "python_version < \"3.11\""},
-    {version = ">=1.23.2", markers = "python_version == \"3.11\""},
     {version = ">=1.26.0", markers = "python_version >= \"3.12\""},
+    {version = ">=1.23.2", markers = "python_version == \"3.11\""},
+    {version = ">=1.22.4", markers = "python_version < \"3.11\""},
 ]
 python-dateutil = ">=2.8.2"
 pytz = ">=2020.1"
@@ -3623,13 +3893,13 @@ typing-extensions = ">=4.6.0,<4.7.0 || >4.7.0"
 
 [[package]]
 name = "pydantic-settings"
-version = "2.8.0"
+version = "2.8.1"
 description = "Settings management using Pydantic"
 optional = false
 python-versions = ">=3.8"
 files = [
-    {file = "pydantic_settings-2.8.0-py3-none-any.whl", hash = "sha256:c782c7dc3fb40e97b238e713c25d26f64314aece2e91abcff592fcac15f71820"},
-    {file = "pydantic_settings-2.8.0.tar.gz", hash = "sha256:88e2ca28f6e68ea102c99c3c401d6c9078e68a5df600e97b43891c34e089500a"},
+    {file = "pydantic_settings-2.8.1-py3-none-any.whl", hash = "sha256:81942d5ac3d905f7f3ee1a70df5dfb62d5569c12f51a5a647defc1c3d9ee2e9c"},
+    {file = "pydantic_settings-2.8.1.tar.gz", hash = "sha256:d5c663dfbe9db9d5e1c646b2e161da12f0d734d422ee56f567d0ea2cee4e8585"},
 ]
 
 [package.dependencies]
@@ -3660,6 +3930,21 @@ numpy = ">=1.16.4"
 carto = ["pydeck-carto"]
 jupyter = ["ipykernel (>=5.1.2)", "ipython (>=5.8.0)", "ipywidgets (>=7,<8)", "traitlets (>=4.3.2)"]
 
+[[package]]
+name = "pydyf"
+version = "0.11.0"
+description = "A low-level PDF generator."
+optional = true
+python-versions = ">=3.8"
+files = [
+    {file = "pydyf-0.11.0-py3-none-any.whl", hash = "sha256:0aaf9e2ebbe786ec7a78ec3fbffa4cdcecde53fd6f563221d53c6bc1328848a3"},
+    {file = "pydyf-0.11.0.tar.gz", hash = "sha256:394dddf619cca9d0c55715e3c55ea121a9bf9cbc780cdc1201a2427917b86b64"},
+]
+
+[package.extras]
+doc = ["sphinx", "sphinx_rtd_theme"]
+test = ["pillow", "pytest", "ruff"]
+
 [[package]]
 name = "pyee"
 version = "12.1.1"
@@ -3713,6 +3998,21 @@ files = [
     {file = "pypdfium2-4.30.0.tar.gz", hash = "sha256:48b5b7e5566665bc1015b9d69c1ebabe21f6aee468b509531c3c8318eeee2e16"},
 ]
 
+[[package]]
+name = "pyphen"
+version = "0.17.2"
+description = "Pure Python module to hyphenate text"
+optional = true
+python-versions = ">=3.9"
+files = [
+    {file = "pyphen-0.17.2-py3-none-any.whl", hash = "sha256:3a07fb017cb2341e1d9ff31b8634efb1ae4dc4b130468c7c39dd3d32e7c3affd"},
+    {file = "pyphen-0.17.2.tar.gz", hash = "sha256:f60647a9c9b30ec6c59910097af82bc5dd2d36576b918e44148d8b07ef3b4aa3"},
+]
+
+[package.extras]
+doc = ["sphinx", "sphinx_rtd_theme"]
+test = ["pytest", "ruff"]
+
 [[package]]
 name = "pytest"
 version = "8.3.4"
@@ -3809,7 +4109,7 @@ files = [
 name = "python-pptx"
 version = "1.0.2"
 description = "Create, read, and update PowerPoint 2007+ (.pptx) files."
-optional = false
+optional = true
 python-versions = ">=3.8"
 files = [
     {file = "python_pptx-1.0.2-py3-none-any.whl", hash = "sha256:160838e0b8565a8b1f67947675886e9fea18aa5e795db7ae531606d68e785cba"},
@@ -4641,13 +4941,13 @@ win32 = ["pywin32"]
 
 [[package]]
 name = "setuptools"
-version = "75.8.1"
+version = "75.8.2"
 description = "Easily download, build, install, upgrade, and uninstall Python packages"
 optional = false
 python-versions = ">=3.9"
 files = [
-    {file = "setuptools-75.8.1-py3-none-any.whl", hash = "sha256:3bc32c0b84c643299ca94e77f834730f126efd621de0cc1de64119e0e17dab1f"},
-    {file = "setuptools-75.8.1.tar.gz", hash = "sha256:65fb779a8f28895242923582eadca2337285f0891c2c9e160754df917c3d2530"},
+    {file = "setuptools-75.8.2-py3-none-any.whl", hash = "sha256:558e47c15f1811c1fa7adbd0096669bf76c1d3f433f58324df69f3f5ecac4e8f"},
+    {file = "setuptools-75.8.2.tar.gz", hash = "sha256:4880473a969e5f23f2a2be3646b2dfd84af9028716d398e46192f84bc36900d2"},
 ]
 
 [package.extras]
@@ -4724,13 +5024,13 @@ tests = ["cython", "littleutils", "pygments", "pytest", "typeguard"]
 
 [[package]]
 name = "starlette"
-version = "0.45.3"
+version = "0.46.0"
 description = "The little ASGI library that shines."
 optional = false
 python-versions = ">=3.9"
 files = [
-    {file = "starlette-0.45.3-py3-none-any.whl", hash = "sha256:dfb6d332576f136ec740296c7e8bb8c8a7125044e7c6da30744718880cdd059d"},
-    {file = "starlette-0.45.3.tar.gz", hash = "sha256:2cbcba2a75806f8a41c722141486f37c28e30a0921c5f6fe4346cb0dcee1302f"},
+    {file = "starlette-0.46.0-py3-none-any.whl", hash = "sha256:913f0798bd90ba90a9156383bcf1350a17d6259451d0d8ee27fc0cf2db609038"},
+    {file = "starlette-0.46.0.tar.gz", hash = "sha256:b359e4567456b28d473d0193f34c0de0ed49710d75ef183a74a5ce0499324f50"},
 ]
 
 [package.dependencies]
@@ -4894,6 +5194,24 @@ webencodings = ">=0.4"
 doc = ["sphinx", "sphinx_rtd_theme"]
 test = ["pytest", "ruff"]
 
+[[package]]
+name = "tinyhtml5"
+version = "2.0.0"
+description = "HTML parser based on the WHATWG HTML specification"
+optional = true
+python-versions = ">=3.9"
+files = [
+    {file = "tinyhtml5-2.0.0-py3-none-any.whl", hash = "sha256:13683277c5b176d070f82d099d977194b7a1e26815b016114f581a74bbfbf47e"},
+    {file = "tinyhtml5-2.0.0.tar.gz", hash = "sha256:086f998833da24c300c414d9fe81d9b368fd04cb9d2596a008421cbc705fcfcc"},
+]
+
+[package.dependencies]
+webencodings = ">=0.5.1"
+
+[package.extras]
+doc = ["sphinx", "sphinx_rtd_theme"]
+test = ["pytest", "ruff"]
+
 [[package]]
 name = "tokenizers"
 version = "0.21.0"
@@ -5314,6 +5632,31 @@ files = [
     {file = "wcwidth-0.2.13.tar.gz", hash = "sha256:72ea0c06399eb286d978fdedb6923a9eb47e1c486ce63e9b4e64fc18303972b5"},
 ]
 
+[[package]]
+name = "weasyprint"
+version = "63.1"
+description = "The Awesome Document Factory"
+optional = true
+python-versions = ">=3.9"
+files = [
+    {file = "weasyprint-63.1-py3-none-any.whl", hash = "sha256:9d0319fe3ba553c9a77dc43a2d35b64a70c2b8809ad55a139a214803fde62bce"},
+    {file = "weasyprint-63.1.tar.gz", hash = "sha256:cb424e63e8dd3f14195bfe5f203527646aa40a2f00ac819f9d39b8304cec0044"},
+]
+
+[package.dependencies]
+cffi = ">=0.6"
+cssselect2 = ">=0.1"
+fonttools = {version = ">=4.0.0", extras = ["woff"]}
+Pillow = ">=9.1.0"
+pydyf = ">=0.11.0"
+Pyphen = ">=0.9.1"
+tinycss2 = ">=1.4.0"
+tinyhtml5 = ">=2.0.0b1"
+
+[package.extras]
+doc = ["sphinx", "sphinx_rtd_theme"]
+test = ["pytest", "ruff"]
+
 [[package]]
 name = "webcolors"
 version = "24.11.1"
@@ -5445,7 +5788,7 @@ files = [
 name = "xlsxwriter"
 version = "3.2.2"
 description = "A Python module for creating Excel XLSX files."
-optional = false
+optional = true
 python-versions = ">=3.6"
 files = [
     {file = "XlsxWriter-3.2.2-py3-none-any.whl", hash = "sha256:272ce861e7fa5e82a4a6ebc24511f2cb952fde3461f6c6e1a1e81d3272db1471"},
@@ -5680,7 +6023,93 @@ idna = ">=2.0"
 multidict = ">=4.0"
 propcache = ">=0.2.0"
 
+[[package]]
+name = "zopfli"
+version = "0.2.3.post1"
+description = "Zopfli module for python"
+optional = true
+python-versions = ">=3.8"
+files = [
+    {file = "zopfli-0.2.3.post1-cp310-cp310-macosx_10_9_universal2.whl", hash = "sha256:e0137dd64a493ba6a4be37405cfd6febe650a98cc1e9dca8f6b8c63b1db11b41"},
+    {file = "zopfli-0.2.3.post1-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:aa588b21044f8a74e423d8c8a4c7fc9988501878aacced793467010039c50734"},
+    {file = "zopfli-0.2.3.post1-cp310-cp310-manylinux_2_12_i686.manylinux2010_i686.whl", hash = "sha256:9f4a7ec2770e6af05f5a02733fd3900f30a9cd58e5d6d3727e14c5bcd6e7d587"},
+    {file = "zopfli-0.2.3.post1-cp310-cp310-manylinux_2_12_x86_64.manylinux2010_x86_64.whl", hash = "sha256:f7d69c1a7168ad0e9cb864e8663acb232986a0c9c9cb9801f56bf6214f53a54d"},
+    {file = "zopfli-0.2.3.post1-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:6c2d2bc8129707e34c51f9352c4636ca313b52350bbb7e04637c46c1818a2a70"},
+    {file = "zopfli-0.2.3.post1-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:39e576f93576c5c223b41d9c780bbb91fd6db4babf3223d2a4fe7bf568e2b5a8"},
+    {file = "zopfli-0.2.3.post1-cp310-cp310-musllinux_1_2_i686.whl", hash = "sha256:cbe6df25807227519debd1a57ab236f5f6bad441500e85b13903e51f93a43214"},
+    {file = "zopfli-0.2.3.post1-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:7cce242b5df12b2b172489daf19c32e5577dd2fac659eb4b17f6a6efb446fd5c"},
+    {file = "zopfli-0.2.3.post1-cp310-cp310-win32.whl", hash = "sha256:f815fcc2b2a457977724bad97fb4854022980f51ce7b136925e336b530545ae1"},
+    {file = "zopfli-0.2.3.post1-cp310-cp310-win_amd64.whl", hash = "sha256:0cc20b02a9531559945324c38302fd4ba763311632d0ec8a1a0aa9c10ea363e6"},
+    {file = "zopfli-0.2.3.post1-cp311-cp311-macosx_10_9_universal2.whl", hash = "sha256:518f1f4ed35dd69ce06b552f84e6d081f07c552b4c661c5312d950a0b764a58a"},
+    {file = "zopfli-0.2.3.post1-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:615a8ac9dda265e9cc38b2a76c3142e4a9f30fea4a79c85f670850783bc6feb4"},
+    {file = "zopfli-0.2.3.post1-cp311-cp311-manylinux_2_12_i686.manylinux2010_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:a82fc2dbebe6eb908b9c665e71496f8525c1bc4d2e3a7a7722ef2b128b6227c8"},
+    {file = "zopfli-0.2.3.post1-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:37d011e92f7b9622742c905fdbed9920a1d0361df84142807ea2a528419dea7f"},
+    {file = "zopfli-0.2.3.post1-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:e63d558847166543c2c9789e6f985400a520b7eacc4b99181668b2c3aeadd352"},
+    {file = "zopfli-0.2.3.post1-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:60db20f06c3d4c5934b16cfa62a2cc5c3f0686bffe0071ed7804d3c31ab1a04e"},
+    {file = "zopfli-0.2.3.post1-cp311-cp311-musllinux_1_2_i686.whl", hash = "sha256:716cdbfc57bfd3d3e31a58e6246e8190e6849b7dbb7c4ce39ef8bbf0edb8f6d5"},
+    {file = "zopfli-0.2.3.post1-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:3a89277ed5f8c0fb2d0b46d669aa0633123aa7381f1f6118c12f15e0fb48f8ca"},
+    {file = "zopfli-0.2.3.post1-cp311-cp311-win32.whl", hash = "sha256:75a26a2307b10745a83b660c404416e984ee6fca515ec7f0765f69af3ce08072"},
+    {file = "zopfli-0.2.3.post1-cp311-cp311-win_amd64.whl", hash = "sha256:81c341d9bb87a6dbbb0d45d6e272aca80c7c97b4b210f9b6e233bf8b87242f29"},
+    {file = "zopfli-0.2.3.post1-cp312-cp312-macosx_10_13_universal2.whl", hash = "sha256:3f0197b6aa6eb3086ae9e66d6dd86c4d502b6c68b0ec490496348ae8c05ecaef"},
+    {file = "zopfli-0.2.3.post1-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:5fcfc0dc2761e4fcc15ad5d273b4d58c2e8e059d3214a7390d4d3c8e2aee644e"},
+    {file = "zopfli-0.2.3.post1-cp312-cp312-manylinux_2_12_i686.manylinux2010_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:cac2b37ab21c2b36a10b685b1893ebd6b0f83ae26004838ac817680881576567"},
+    {file = "zopfli-0.2.3.post1-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:8d5ab297d660b75c159190ce6d73035502310e40fd35170aed7d1a1aea7ddd65"},
+    {file = "zopfli-0.2.3.post1-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:9ba214f4f45bec195ee8559651154d3ac2932470b9d91c5715fc29c013349f8c"},
+    {file = "zopfli-0.2.3.post1-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:c1e0ed5d84ffa2d677cc9582fc01e61dab2e7ef8b8996e055f0a76167b1b94df"},
+    {file = "zopfli-0.2.3.post1-cp312-cp312-musllinux_1_2_i686.whl", hash = "sha256:bfa1eb759e07d8b7aa7a310a2bc535e127ee70addf90dc8d4b946b593c3e51a8"},
+    {file = "zopfli-0.2.3.post1-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:cd2c002f160502608dcc822ed2441a0f4509c52e86fcfd1a09e937278ed1ca14"},
+    {file = "zopfli-0.2.3.post1-cp312-cp312-win32.whl", hash = "sha256:7be5cc6732eb7b4df17305d8a7b293223f934a31783a874a01164703bc1be6cd"},
+    {file = "zopfli-0.2.3.post1-cp312-cp312-win_amd64.whl", hash = "sha256:4e50ffac74842c1c1018b9b73875a0d0a877c066ab06bf7cccbaa84af97e754f"},
+    {file = "zopfli-0.2.3.post1-cp313-cp313-macosx_10_13_universal2.whl", hash = "sha256:ecb7572df5372abce8073df078207d9d1749f20b8b136089916a4a0868d56051"},
+    {file = "zopfli-0.2.3.post1-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:a1cf720896d2ce998bc8e051d4b4ce0d8bec007aab6243102e8e1d22a0b2fb3f"},
+    {file = "zopfli-0.2.3.post1-cp313-cp313-manylinux_2_12_i686.manylinux2010_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:5aad740b4d4fcbaaae4887823925166ffd062db3b248b3f432198fc287381d1a"},
+    {file = "zopfli-0.2.3.post1-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:6617fb10f9e4393b331941861d73afb119cd847e88e4974bdbe8068ceef3f73f"},
+    {file = "zopfli-0.2.3.post1-cp313-cp313-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:a53b18797cdef27e019db595d66c4b077325afe2fd62145953275f53d84ce40c"},
+    {file = "zopfli-0.2.3.post1-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:b78008a69300d929ca2efeffec951b64a312e9a811e265ea4a907ab546d79fa6"},
+    {file = "zopfli-0.2.3.post1-cp313-cp313-musllinux_1_2_i686.whl", hash = "sha256:0aa5f90d6298bda02a95bc8dc8c3c19004d5a4e44bda00b67ca7431d857b4b54"},
+    {file = "zopfli-0.2.3.post1-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:2768c877f76c8a0e7519b1c86c93757f3c01492ddde55751e9988afb7eff64e1"},
+    {file = "zopfli-0.2.3.post1-cp313-cp313-win32.whl", hash = "sha256:71390dbd3fbf6ebea9a5d85ffed8c26ee1453ee09248e9b88486e30e0397b775"},
+    {file = "zopfli-0.2.3.post1-cp313-cp313-win_amd64.whl", hash = "sha256:a86eb88e06bd87e1fff31dac878965c26b0c26db59ddcf78bb0379a954b120de"},
+    {file = "zopfli-0.2.3.post1-cp38-cp38-macosx_10_9_universal2.whl", hash = "sha256:3827170de28faf144992d3d4dcf8f3998fe3c8a6a6f4a08f1d42c2ec6119d2bb"},
+    {file = "zopfli-0.2.3.post1-cp38-cp38-macosx_10_9_x86_64.whl", hash = "sha256:b0ec13f352ea5ae0fc91f98a48540512eed0767d0ec4f7f3cb92d92797983d18"},
+    {file = "zopfli-0.2.3.post1-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:5f272186e03ad55e7af09ab78055535c201b1a0bcc2944edb1768298d9c483a4"},
+    {file = "zopfli-0.2.3.post1-cp38-cp38-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:29ea74e72ffa6e291b8c6f2504ce6c146b4fe990c724c1450eb8e4c27fd31431"},
+    {file = "zopfli-0.2.3.post1-cp38-cp38-manylinux_2_5_x86_64.manylinux1_x86_64.whl", hash = "sha256:eb45a34f23da4f8bc712b6376ca5396914b0b7c09adbb001dad964eb7f3132f8"},
+    {file = "zopfli-0.2.3.post1-cp38-cp38-musllinux_1_2_aarch64.whl", hash = "sha256:6482db9876c68faac2d20a96b566ffbf65ddaadd97b222e4e73641f4f8722fc4"},
+    {file = "zopfli-0.2.3.post1-cp38-cp38-musllinux_1_2_i686.whl", hash = "sha256:95a260cafd56b8fffa679918937401c80bb38e1681c448b988022e4c3610965d"},
+    {file = "zopfli-0.2.3.post1-cp38-cp38-musllinux_1_2_x86_64.whl", hash = "sha256:676919fba7311125244eb0c4393679ac5fe856e5864a15d122bd815205369fa0"},
+    {file = "zopfli-0.2.3.post1-cp38-cp38-win32.whl", hash = "sha256:b9026a21b6d41eb0e2e63f5bc1242c3fcc43ecb770963cda99a4307863dac12e"},
+    {file = "zopfli-0.2.3.post1-cp38-cp38-win_amd64.whl", hash = "sha256:3c163911f8bad94b3e1db0a572e7c28ba681a0c91d0002ea1e4fa9264c21ef17"},
+    {file = "zopfli-0.2.3.post1-cp39-cp39-macosx_10_9_universal2.whl", hash = "sha256:b05296e8bc88c92e2b21e0a9bae4740c1551ee613c1d93a51fd28a7a0b2b6fbb"},
+    {file = "zopfli-0.2.3.post1-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:f12000a6accdd4bf0a3fa6eaa1b1c7a7bc80af0a2edf3f89d770d3dcce1d0e22"},
+    {file = "zopfli-0.2.3.post1-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:a241a68581d34d67b40c425cce3d1fd211c092f99d9250947824ccba9f491949"},
+    {file = "zopfli-0.2.3.post1-cp39-cp39-manylinux_2_5_i686.manylinux1_i686.whl", hash = "sha256:3657e416ffb8f31d9d3424af12122bb251befae109f2e271d87d825c92fc5b7b"},
+    {file = "zopfli-0.2.3.post1-cp39-cp39-manylinux_2_5_x86_64.manylinux1_x86_64.whl", hash = "sha256:4915a41375bdee4db749ecd07d985a0486eb688a6619f713b7bf6fbfd145e960"},
+    {file = "zopfli-0.2.3.post1-cp39-cp39-musllinux_1_2_aarch64.whl", hash = "sha256:bbe429fc50686bb2a2608a30843e36fbaa123462a5284f136c7d9e0145220bfd"},
+    {file = "zopfli-0.2.3.post1-cp39-cp39-musllinux_1_2_i686.whl", hash = "sha256:2345e713260a350bea0b01a816a469ea356bc2d63d009a0d777691ecbbcf7493"},
+    {file = "zopfli-0.2.3.post1-cp39-cp39-musllinux_1_2_x86_64.whl", hash = "sha256:fc39f5c27f962ec8660d8d20c24762431131b5d8c672b44b0a54cf2b5bcde9b9"},
+    {file = "zopfli-0.2.3.post1-cp39-cp39-win32.whl", hash = "sha256:9a6aec38a989bad7ddd1ef53f1265699e49e294d08231b5313d61293f3cd6237"},
+    {file = "zopfli-0.2.3.post1-cp39-cp39-win_amd64.whl", hash = "sha256:b3df42f52502438ee973042cc551877d24619fa1cd38ef7b7e9ac74200daca8b"},
+    {file = "zopfli-0.2.3.post1-pp310-pypy310_pp73-macosx_10_15_x86_64.whl", hash = "sha256:4c1226a7e2c7105ac31503a9bb97454743f55d88164d6d46bc138051b77f609b"},
+    {file = "zopfli-0.2.3.post1-pp310-pypy310_pp73-manylinux_2_12_i686.manylinux2010_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:48dba9251060289101343110ab47c0756f66f809bb4d1ddbb6d5c7e7752115c5"},
+    {file = "zopfli-0.2.3.post1-pp310-pypy310_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:89899641d4de97dbad8e0cde690040d078b6aea04066dacaab98e0b5a23573f2"},
+    {file = "zopfli-0.2.3.post1-pp310-pypy310_pp73-win_amd64.whl", hash = "sha256:3654bfc927bc478b1c3f3ff5056ed7b20a1a37fa108ca503256d0a699c03bbb1"},
+    {file = "zopfli-0.2.3.post1-pp38-pypy38_pp73-macosx_10_9_x86_64.whl", hash = "sha256:c4278d1873ce6e803e5d4f8d702fd3026bd67fca744aa98881324d1157ddf748"},
+    {file = "zopfli-0.2.3.post1-pp38-pypy38_pp73-win_amd64.whl", hash = "sha256:1d8cc06605519e82b16df090e17cb3990d1158861b2872c3117f1168777b81e4"},
+    {file = "zopfli-0.2.3.post1-pp39-pypy39_pp73-macosx_10_15_x86_64.whl", hash = "sha256:1f990634fd5c5c8ced8edddd8bd45fab565123b4194d6841e01811292650acae"},
+    {file = "zopfli-0.2.3.post1-pp39-pypy39_pp73-manylinux_2_12_i686.manylinux2010_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:91a2327a4d7e77471fa4fbb26991c6de4a738c6fc6a33e09bb25f56a870a4b7b"},
+    {file = "zopfli-0.2.3.post1-pp39-pypy39_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:8fbe5bcf10d01aab3513550f284c09fef32f342b36f56bfae2120a9c4d12c130"},
+    {file = "zopfli-0.2.3.post1-pp39-pypy39_pp73-win_amd64.whl", hash = "sha256:34a99592f3d9eb6f737616b5bd74b48a589fdb3cb59a01a50d636ea81d6af272"},
+    {file = "zopfli-0.2.3.post1.tar.gz", hash = "sha256:96484dc0f48be1c5d7ae9f38ed1ce41e3675fd506b27c11a6607f14b49101e99"},
+]
+
+[package.extras]
+test = ["pytest"]
+
+[extras]
+full = ["ebooklib", "mammoth", "openpyxl", "python-pptx", "weasyprint"]
+
 [metadata]
 lock-version = "2.0"
 python-versions = "^3.10"
-content-hash = "70d4b1cc842480888768dbad78c97cdedeec150ccb3433e46c030327f2eb7819"
+content-hash = "7d16a2f72bebb9a040c49a446dc9951b9e50bc88a007e2f646d0786206cdfde4"

From d16455f36c2e55ed99a57efd7bea332bda15b860 Mon Sep 17 00:00:00 2001
From: Vik Paruchuri <vik.paruchuri@gmail.com>
Date: Fri, 28 Feb 2025 15:34:34 -0500
Subject: [PATCH 43/46] Bump surya version

---
 marker/converters/pdf.py                |   8 +-
 marker/processors/llm/llm_inlinemath.py | 238 ++++++++++++------------
 marker/processors/llm/llm_mathblock.py  | 164 ++++++++++++++++
 marker/processors/llm/llm_text.py       | 168 -----------------
 marker/scripts/streamlit_app.py         |   2 +
 poetry.lock                             |  55 +-----
 pyproject.toml                          |   2 +-
 tests/processors/test_inline_math.py    |   4 +-
 tests/processors/test_llm_processors.py |   2 +-
 9 files changed, 305 insertions(+), 338 deletions(-)
 create mode 100644 marker/processors/llm/llm_mathblock.py
 delete mode 100644 marker/processors/llm/llm_text.py

diff --git a/marker/converters/pdf.py b/marker/converters/pdf.py
index abab04ed..ccd03f03 100644
--- a/marker/converters/pdf.py
+++ b/marker/converters/pdf.py
@@ -27,7 +27,7 @@
 from marker.processors.llm.llm_form import LLMFormProcessor
 from marker.processors.llm.llm_image_description import LLMImageDescriptionProcessor
 from marker.processors.llm.llm_table import LLMTableProcessor
-from marker.processors.llm.llm_text import LLMTextProcessor
+from marker.processors.llm.llm_inlinemath import LLMInlineMathLinesProcessor
 from marker.processors.page_header import PageHeaderProcessor
 from marker.processors.reference import ReferenceProcessor
 from marker.processors.sectionheader import SectionHeaderProcessor
@@ -43,7 +43,7 @@
 from marker.processors.order import OrderProcessor
 from marker.services.gemini import GoogleGeminiService
 from marker.processors.line_merge import LineMergeProcessor
-from marker.processors.llm.llm_inlinemath import LLMInlineMathProcessor
+from marker.processors.llm.llm_mathblock import LLMMathBlockProcessor
 
 
 class PdfConverter(BaseConverter):
@@ -79,12 +79,12 @@ class PdfConverter(BaseConverter):
         LLMTableMergeProcessor,
         LLMFormProcessor,
         TextProcessor,
-        LLMTextProcessor,
+        LLMInlineMathLinesProcessor,
         LLMComplexRegionProcessor,
         LLMImageDescriptionProcessor,
         LLMEquationProcessor,
         LLMHandwritingProcessor,
-        LLMInlineMathProcessor,
+        LLMMathBlockProcessor,
         ReferenceProcessor,
         DebugProcessor,
     )
diff --git a/marker/processors/llm/llm_inlinemath.py b/marker/processors/llm/llm_inlinemath.py
index 0ab43907..8c32adc4 100644
--- a/marker/processors/llm/llm_inlinemath.py
+++ b/marker/processors/llm/llm_inlinemath.py
@@ -1,164 +1,168 @@
 import json
-from concurrent.futures import ThreadPoolExecutor, as_completed
 from typing import List, Tuple, Annotated
 
 from pydantic import BaseModel
-from tqdm import tqdm
+from PIL import Image
 
-from marker.processors.llm import BaseLLMComplexBlockProcessor
+from marker.processors.llm import BaseLLMSimpleBlockProcessor, PromptData, BlockData
 
+from marker.processors.util import add_math_spans_to_line
 from marker.schema import BlockTypes
-from marker.schema.blocks import Block, InlineMath
+from marker.schema.blocks import Block
 from marker.schema.document import Document
-from marker.schema.groups import PageGroup
+from marker.schema.text import Line
 
 
-class LLMInlineMathProcessor(BaseLLMComplexBlockProcessor):
-    redo_inline_math: Annotated[
-        bool,
-        "If True, the inline math will be re-done, otherwise it will be left as is."
-    ] = False
-    inlinemath_min_ratio: Annotated[
-        float,
-        "If more than this ratio of blocks are inlinemath blocks, assume everything has math."
-    ] = 0.4
+class LLMInlineMathLinesProcessor(BaseLLMSimpleBlockProcessor):
+    math_line_batch_size: Annotated[
+        int,
+        "The number of math lines to batch together.",
+    ] = 10
 
-    block_types = (BlockTypes.TextInlineMath,) # Primary block type
-    additional_block_types = (BlockTypes.Text, BlockTypes.Caption, BlockTypes.SectionHeader, BlockTypes.Footnote) # Seconday, can also contain math
-
-    text_math_rewriting_prompt = """You are a text correction expert specializing in accurately reproducing text from images.
+    block_types = (BlockTypes.Line,)
+    image_remove_blocks = (BlockTypes.Equation,)
+    text_math_rewriting_prompt = r"""You are a text correction expert specializing in accurately reproducing text from images.
 You will receive an image of a text block and a set of extracted lines corresponding to the text in the image.
-Your task is to correct any errors in the extracted block, including math, formatting, and other inaccuracies, and output the corrected block in html format.  Stay as faithful to the original text as possible.
+Your task is to correct any errors in the extracted lines, including math, formatting, and other inaccuracies, and output the corrected lines in a JSON format.
+
+The number of output lines MUST match the number of input lines.  Stay as faithful to the original text as possible.
 
 **Instructions:**
 
 1. Carefully examine the provided text block image .
-2. Analyze the text that has been extracted from the block.
-3. Compare the extracted text to the corresponding text in the image.
-4. Write a short analysis of the text block, including any errors you see in the extracted text.
-5. If there are no errors in any of the extracted text, output "No corrections needed".
-6. Correct any errors in the extracted text, including:
-    * Inline math: Ensure all mathematical expressions are correctly formatted and rendered.  Surround them with <math>...</math> tags.  The math expressions should be rendered in simple, concise, KaTeX-compatible LaTeX.  Do not use $ or $$ as delimiters.
-      * If a math expression is not in LaTeX format, convert it to LaTeX format, and surround it with <math>...</math> tags.
-    * Formatting: Maintain consistent formatting with the text block image, including spacing, indentation, subscripts/superscripts, and special characters.  Use the <h1>, <h2>, <h3>, <h4>, <i>, <b>, <sup>, <sub>, and <span> tags to format the text as needed.
+2. Analyze the extracted lines.
+3. Write a short analysis comparing the extracted lines to the image.
+4. For each extracted line, compare it to the corresponding line in the image.
+5. Correct any errors in the extracted line, including:
+    * Inline math: Ensure all mathematical expressions are correctly formatted and rendered.  Use the `<math>` and `</math>` tags to surround inline math properly.  Make sure the opening and closing tags appear in pairs, on the same line.  The math should be written in simple, concise, KaTeX-compatible LaTeX.  Do not use $ or $$ as delimiters.
+    * Formatting: Maintain consistent formatting with the text block image, including spacing, indentation, subscripts/superscripts, and special characters.  Use the `<i>`, `<b>`, `<sup>`, `<sub>`, and `<span>` tags to format the text as needed.
     * Other inaccuracies:  If the image is handwritten then you may correct any spelling errors, or other discrepancies.
-    * Ensure lines wrap properly, and that newlines are not in the middle of sentences.
-7. Do not remove any formatting i.e bold, italics, math, superscripts, subscripts, etc from the extracted text unless it is necessary to correct an error.
-8. Output the corrected text in html format, as shown in the example below.  Only use the h1, h2, h3, h4, p, math, br, a, i, b, sup, sub, and span tags.
+6. Do not remove any formatting i.e bold, italics, math, superscripts, subscripts, etc from the extracted lines unless it is necessary to correct an error.  The formatting 
+7. The number of corrected lines in the output MUST equal the number of extracted lines provided in the input. Do not add or remove lines.  There are exactly {line_count} input lines.
+8. Output the corrected lines in JSON format, as shown in the example below.  Each line should be in HTML format. Only use the math, br, a, i, b, sup, sub, and span tags.
 9. You absolutely cannot remove any <a href='#...'>...</a> tags, those are extremely important for references and are coming directly from the document, you MUST always preserve them.
 
 **Example:**
 
 Input:
-```html
-Adversarial training (AT) <a href='#page-9-1'>[23]</a>, which aims to minimize the model's risk under the worst-case perturbations, 
-is currently the most effective approach for improving the robustness of deep neural networks. For a given neural network f(x, w) 
-with parameters w, the optimization objective of AT can be formulated as follows:
+```
+{
+ "extracted_lines": [
+  "Adversarial training (AT) <a href='#page-9-1'>[23]</a>, which aims to minimize\n",
+  "the model's risk under the worst-case perturbations, is cur-\n",
+  "rently the most effective approach for improving the robust-\n",
+  "ness of deep neural networks. For a given neural network\n",
+  "f(x, w) with parameters w, the optimization objective of\n",
+  "AT can be formulated as follows:\n"
+ ]
+}
 ```
 
 Output:
-analysis: The inline math is not in LaTeX format and is not surrounded by <math>...</math> tags.
-```html
-Adversarial training <i>(AT)</i> <a href='#page-9-1'>[23]</a>, which aims to minimize the model's risk under the worst-case perturbations, is currently the most effective approach for improving the robustness of deep neural networks. For a given neural network <math>f(x, w)</math> with parameters <math>w</math>, the optimization objective of AT can be formulated as follows:
+analysis: The inline math in the lines is not in LaTeX format and is not surrounded by <math>...</math> tags.
+```json
+{
+ "corrected_lines": [
+  "Adversarial training (AT) <a href='#page-9-1'>[23]</a>, which aims to minimize\n",
+  "the model's risk under the worst-case perturbations, is cur-\n",
+  "rently the most effective approach for improving the robust-\n",
+  "ness of deep neural networks. For a given neural network\n",
+  "<math>f(x, w)</math> with parameters <math>w</math>, the optimization objective of\n",
+  "AT can be formulated as follows:\n"
+ ]
+}
 ```
 
 **Input:**
-```html
-{extracted_html}
+```json
+{extracted_lines}
 ```
 """
 
-    def rewrite_blocks(self, document: Document):
-        if not self.redo_inline_math:
-            return
-
-        # Get inline math blocks
-        inline_blocks: List[InlineMath] = [
-            (page, block)
-            for page in document.pages
-            for block in page.contained_blocks(document, self.block_types)
-        ]
-
-        # Get other blocks with detected math in them
-        detected_blocks = [
-            (page, block)
-            for page in document.pages
-            for block in page.contained_blocks(document, (BlockTypes.Text, BlockTypes.Caption, BlockTypes.SectionHeader, BlockTypes.Footnote, BlockTypes.ListItem))
-            if any([b.formats and "math" in b.formats for b in block.contained_blocks(document, (BlockTypes.Line,))])
-        ]
-
-        # If a page has enough math blocks, assume all blocks can contain math
-        additional_text_blocks = []
+    def inference_blocks(self, document: Document) -> List[List[BlockData]]:
+        blocks = []
         for page in document.pages:
-            # Check for inline math blocks
-            page_inlinemath_blocks = [im for im in inline_blocks if im[0].page_id == page.page_id]
-            page_detected_blocks = [db for db in detected_blocks if db[0].page_id == page.page_id]
-            math_block_count = len(page_inlinemath_blocks) + len(page_detected_blocks)
-
-            # Find all potential blocks
-            additional_blocks = page.contained_blocks(document, self.additional_block_types + self.block_types)
-
-            # Check if the ratio of math blocks to additional blocks is high enough
-            if math_block_count / max(1, len(additional_blocks)) < self.inlinemath_min_ratio:
-                continue
-
-            for b in additional_blocks:
-                if b not in detected_blocks and b not in inline_blocks:
-                    additional_text_blocks.append((page, b))
-
-        inference_blocks = inline_blocks + detected_blocks + additional_text_blocks
+            for block in page.contained_blocks(document, self.block_types):
+                if block.formats and "math" in block.formats:
+                    blocks.append({
+                        "page": page,
+                        "block": block
+                    })
 
-        # Don't show progress if there are no blocks to process
-        total_blocks = len(inference_blocks)
-        if total_blocks == 0:
-            return
-
-        pbar = tqdm(desc=f"{self.__class__.__name__} running", disable=self.disable_tqdm)
-        with ThreadPoolExecutor(max_workers=self.max_concurrency) as executor:
-            for future in as_completed([
-                executor.submit(self.process_rewriting, document, b[0], b[1])
-                for b in inference_blocks
-            ]):
-                future.result()  # Raise exceptions if any occurred
-                pbar.update(1)
-
-        pbar.close()
 
-    def get_block_text(self, block: Block, document: Document) -> str:
-        html = block.render(document).html
-        return html
+        out_blocks = []
+        for i in range(0, len(blocks), self.math_line_batch_size):
+            batch = blocks[i:i + self.math_line_batch_size]
+            out_blocks.append(batch)
+        return out_blocks
 
     def get_block_lines(self, block: Block, document: Document) -> Tuple[list, list]:
         text_lines = block.contained_blocks(document, (BlockTypes.Line,))
         extracted_lines = [line.formatted_text(document) for line in text_lines]
         return text_lines, extracted_lines
 
-    def process_rewriting(self, document: Document, page: PageGroup, block: Block):
-        block_text = self.get_block_text(block, document)
-        prompt = self.text_math_rewriting_prompt.replace("{extracted_html}", block_text)
-
-        image = self.extract_image(document, block)
-        response = self.llm_service(prompt, image, block, LLMTextSchema)
-
-        if not response or "corrected_html" not in response:
-            block.update_metadata(llm_error_count=1)
-            return
-
-        corrected_html = response["corrected_html"]
-        if not corrected_html:
-            block.update_metadata(llm_error_count=1)
-            return
-
-        # Block is fine
-        if "no corrections needed" in corrected_html.lower():
+    def combine_images(self, images: List[Image.Image]):
+        widths, heights = zip(*(i.size for i in images))
+        total_width = max(widths)
+        total_height = sum(heights) + 5 * len(images)
+
+        new_im = Image.new('RGB', (total_width, total_height), (255, 255, 255))
+
+        y_offset = 0
+        for im in images:
+            new_im.paste(im, (0, y_offset))
+            y_offset += im.size[1] + 5
+
+        return new_im
+
+    def block_prompts(self, document: Document) -> List[PromptData]:
+        prompt_data = []
+        for block_data in self.inference_blocks(document):
+            blocks: List[Line] = [b["block"] for b in block_data]
+            pages = [b["page"] for b in block_data]
+            block_lines = [block.formatted_text(document) for block in blocks]
+
+            prompt = (
+                self.text_math_rewriting_prompt
+                  .replace("{extracted_lines}",json.dumps({"extracted_lines": block_lines}, indent=2))
+                  .replace("{line_count}", str(len(block_lines)))
+            )
+            images = [self.extract_image(document, block, remove_blocks=self.image_remove_blocks) for block in blocks]
+            image = self.combine_images(images)
+
+            prompt_data.append({
+                "prompt": prompt,
+                "image": image,
+                "block": blocks[0],
+                "schema": LLMTextSchema,
+                "page": pages[0],
+                "additional_data": {"blocks": blocks, "pages": pages}
+            })
+        return prompt_data
+
+
+    def rewrite_block(self, response: dict, prompt_data: PromptData, document: Document):
+        blocks = prompt_data["additional_data"]["blocks"]
+        pages = prompt_data["additional_data"]["pages"]
+
+        if not response or "corrected_lines" not in response:
+            blocks[0].update_metadata(llm_error_count=1)
             return
 
-        if len(corrected_html) < len(block_text) * 0.6:
-            block.update_metadata(llm_error_count=1)
+        corrected_lines = response["corrected_lines"]
+        balanced_math = all([line.count("<math") == line.count("</math>") for line in corrected_lines])
+        if any([
+            not corrected_lines,
+            len(corrected_lines) != len(blocks),
+            not balanced_math
+        ]):
+            blocks[0].update_metadata(llm_error_count=1)
             return
 
-        block.html = corrected_html
+        for text_line, page, corrected_text in zip(blocks, pages, corrected_lines):
+            text_line.structure = []
+            add_math_spans_to_line(corrected_text, text_line, page)
 
 class LLMTextSchema(BaseModel):
     analysis: str
-    corrected_html: str
\ No newline at end of file
+    corrected_lines: List[str]
\ No newline at end of file
diff --git a/marker/processors/llm/llm_mathblock.py b/marker/processors/llm/llm_mathblock.py
new file mode 100644
index 00000000..9c191616
--- /dev/null
+++ b/marker/processors/llm/llm_mathblock.py
@@ -0,0 +1,164 @@
+import json
+from concurrent.futures import ThreadPoolExecutor, as_completed
+from typing import List, Tuple, Annotated
+
+from pydantic import BaseModel
+from tqdm import tqdm
+
+from marker.processors.llm import BaseLLMComplexBlockProcessor
+
+from marker.schema import BlockTypes
+from marker.schema.blocks import Block, InlineMath
+from marker.schema.document import Document
+from marker.schema.groups import PageGroup
+
+
+class LLMMathBlockProcessor(BaseLLMComplexBlockProcessor):
+    redo_inline_math: Annotated[
+        bool,
+        "If True, the inline math will be re-done, otherwise it will be left as is."
+    ] = False
+    inlinemath_min_ratio: Annotated[
+        float,
+        "If more than this ratio of blocks are inlinemath blocks, assume everything has math."
+    ] = 0.4
+
+    block_types = (BlockTypes.TextInlineMath,) # Primary block type
+    additional_block_types = (BlockTypes.Text, BlockTypes.Caption, BlockTypes.SectionHeader, BlockTypes.Footnote) # Seconday, can also contain math
+
+    text_math_rewriting_prompt = """You are a text correction expert specializing in accurately reproducing text from images.
+You will receive an image of a text block and a set of extracted lines corresponding to the text in the image.
+Your task is to correct any errors in the extracted block, including math, formatting, and other inaccuracies, and output the corrected block in html format.  Stay as faithful to the original text as possible.
+
+**Instructions:**
+
+1. Carefully examine the provided text block image .
+2. Analyze the text that has been extracted from the block.
+3. Compare the extracted text to the corresponding text in the image.
+4. Write a short analysis of the text block, including any errors you see in the extracted text.
+5. If there are no errors in any of the extracted text, output "No corrections needed".
+6. Correct any errors in the extracted text, including:
+    * Inline math: Ensure all mathematical expressions are correctly formatted and rendered.  Surround them with <math>...</math> tags.  The math expressions should be rendered in simple, concise, KaTeX-compatible LaTeX.  Do not use $ or $$ as delimiters.
+      * If a math expression is not in LaTeX format, convert it to LaTeX format, and surround it with <math>...</math> tags.
+    * Formatting: Maintain consistent formatting with the text block image, including spacing, indentation, subscripts/superscripts, and special characters.  Use the <i>, <b>, <sup>, <sub>, and <span> tags to format the text as needed.
+    * Other inaccuracies:  If the image is handwritten then you may correct any spelling errors, or other discrepancies.
+    * Ensure lines wrap properly, and that newlines are not in the middle of sentences.
+7. Do not remove any formatting i.e bold, italics, math, superscripts, subscripts, etc from the extracted text unless it is necessary to correct an error.
+8. Output the corrected text in html format, as shown in the example below.  Only use the p, math, br, a, i, b, sup, sub, and span tags.
+9. You absolutely cannot remove any <a href='#...'>...</a> tags, those are extremely important for references and are coming directly from the document, you MUST always preserve them.
+
+**Example:**
+
+Input:
+```html
+Adversarial training (AT) <a href='#page-9-1'>[23]</a>, which aims to minimize the model's risk under the worst-case perturbations, 
+is currently the most effective approach for improving the robustness of deep neural networks. For a given neural network f(x, w) 
+with parameters w, the optimization objective of AT can be formulated as follows:
+```
+
+Output:
+analysis: The inline math is not in LaTeX format and is not surrounded by <math>...</math> tags.
+```html
+Adversarial training <i>(AT)</i> <a href='#page-9-1'>[23]</a>, which aims to minimize the model's risk under the worst-case perturbations, is currently the most effective approach for improving the robustness of deep neural networks. For a given neural network <math>f(x, w)</math> with parameters <math>w</math>, the optimization objective of AT can be formulated as follows:
+```
+
+**Input:**
+```html
+{extracted_html}
+```
+"""
+
+    def rewrite_blocks(self, document: Document):
+        if not self.redo_inline_math:
+            return
+
+        # Get inline math blocks
+        inline_blocks: List[InlineMath] = [
+            (page, block)
+            for page in document.pages
+            for block in page.contained_blocks(document, self.block_types)
+        ]
+
+        # Get other blocks with detected math in them
+        detected_blocks = [
+            (page, block)
+            for page in document.pages
+            for block in page.contained_blocks(document, (BlockTypes.Text, BlockTypes.Caption, BlockTypes.SectionHeader, BlockTypes.Footnote, BlockTypes.ListItem))
+            if any([b.formats and "math" in b.formats for b in block.contained_blocks(document, (BlockTypes.Line,))])
+        ]
+
+        # If a page has enough math blocks, assume all blocks can contain math
+        additional_text_blocks = []
+        for page in document.pages:
+            # Check for inline math blocks
+            page_inlinemath_blocks = [im for im in inline_blocks if im[0].page_id == page.page_id]
+            page_detected_blocks = [db for db in detected_blocks if db[0].page_id == page.page_id]
+            math_block_count = len(page_inlinemath_blocks) + len(page_detected_blocks)
+
+            # Find all potential blocks
+            additional_blocks = page.contained_blocks(document, self.additional_block_types + self.block_types)
+
+            # Check if the ratio of math blocks to additional blocks is high enough
+            if math_block_count / max(1, len(additional_blocks)) < self.inlinemath_min_ratio:
+                continue
+
+            for b in additional_blocks:
+                if b not in detected_blocks and b not in inline_blocks:
+                    additional_text_blocks.append((page, b))
+
+        inference_blocks = inline_blocks + detected_blocks + additional_text_blocks
+
+        # Don't show progress if there are no blocks to process
+        total_blocks = len(inference_blocks)
+        if total_blocks == 0:
+            return
+
+        pbar = tqdm(desc=f"{self.__class__.__name__} running", disable=self.disable_tqdm)
+        with ThreadPoolExecutor(max_workers=self.max_concurrency) as executor:
+            for future in as_completed([
+                executor.submit(self.process_rewriting, document, b[0], b[1])
+                for b in inference_blocks
+            ]):
+                future.result()  # Raise exceptions if any occurred
+                pbar.update(1)
+
+        pbar.close()
+
+    def get_block_text(self, block: Block, document: Document) -> str:
+        html = block.render(document).html
+        return html
+
+    def get_block_lines(self, block: Block, document: Document) -> Tuple[list, list]:
+        text_lines = block.contained_blocks(document, (BlockTypes.Line,))
+        extracted_lines = [line.formatted_text(document) for line in text_lines]
+        return text_lines, extracted_lines
+
+    def process_rewriting(self, document: Document, page: PageGroup, block: Block):
+        block_text = self.get_block_text(block, document)
+        prompt = self.text_math_rewriting_prompt.replace("{extracted_html}", block_text)
+
+        image = self.extract_image(document, block)
+        response = self.llm_service(prompt, image, block, LLMTextSchema)
+
+        if not response or "corrected_html" not in response:
+            block.update_metadata(llm_error_count=1)
+            return
+
+        corrected_html = response["corrected_html"]
+        if not corrected_html:
+            block.update_metadata(llm_error_count=1)
+            return
+
+        # Block is fine
+        if "no corrections needed" in corrected_html.lower():
+            return
+
+        if len(corrected_html) < len(block_text) * 0.6:
+            block.update_metadata(llm_error_count=1)
+            return
+
+        block.html = corrected_html
+
+class LLMTextSchema(BaseModel):
+    analysis: str
+    corrected_html: str
\ No newline at end of file
diff --git a/marker/processors/llm/llm_text.py b/marker/processors/llm/llm_text.py
deleted file mode 100644
index a7d85ccf..00000000
--- a/marker/processors/llm/llm_text.py
+++ /dev/null
@@ -1,168 +0,0 @@
-import json
-from typing import List, Tuple, Annotated
-
-from pydantic import BaseModel
-from PIL import Image
-
-from marker.processors.llm import BaseLLMSimpleBlockProcessor, PromptData, BlockData
-
-from marker.processors.util import add_math_spans_to_line
-from marker.schema import BlockTypes
-from marker.schema.blocks import Block
-from marker.schema.document import Document
-from marker.schema.text import Line
-
-
-class LLMTextProcessor(BaseLLMSimpleBlockProcessor):
-    math_line_batch_size: Annotated[
-        int,
-        "The number of math lines to batch together.",
-    ] = 10
-
-    block_types = (BlockTypes.Line,)
-    image_remove_blocks = (BlockTypes.Equation,)
-    text_math_rewriting_prompt = r"""You are a text correction expert specializing in accurately reproducing text from images.
-You will receive an image of a text block and a set of extracted lines corresponding to the text in the image.
-Your task is to correct any errors in the extracted lines, including math, formatting, and other inaccuracies, and output the corrected lines in a JSON format.
-
-The number of output lines MUST match the number of input lines.  Stay as faithful to the original text as possible.
-
-**Instructions:**
-
-1. Carefully examine the provided text block image .
-2. Analyze the extracted lines.
-3. Write a short analysis comparing the extracted lines to the image.
-4. For each extracted line, compare it to the corresponding line in the image.
-5. Correct any errors in the extracted line, including:
-    * Inline math: Ensure all mathematical expressions are correctly formatted and rendered.  Use the `<math>` and `</math>` tags to surround inline math properly.  Make sure the opening and closing tags appear in pairs, on the same line.  The math should be written in simple, concise, KaTeX-compatible LaTeX.  Do not use $ or $$ as delimiters.
-    * Formatting: Maintain consistent formatting with the text block image, including spacing, indentation, subscripts/superscripts, and special characters.  Use the `<i>`, `<b>`, `<sup>`, `<sub>`, and `<span>` tags to format the text as needed.
-    * Other inaccuracies:  If the image is handwritten then you may correct any spelling errors, or other discrepancies.
-6. Do not remove any formatting i.e bold, italics, math, superscripts, subscripts, etc from the extracted lines unless it is necessary to correct an error.  The formatting 
-7. The number of corrected lines in the output MUST equal the number of extracted lines provided in the input. Do not add or remove lines.  There are exactly {line_count} input lines.
-8. Output the corrected lines in JSON format, as shown in the example below.  Each line should be in HTML format. Only use the math, br, a, i, b, sup, sub, and span tags.
-9. You absolutely cannot remove any <a href='#...'>...</a> tags, those are extremely important for references and are coming directly from the document, you MUST always preserve them.
-
-**Example:**
-
-Input:
-```
-{
- "extracted_lines": [
-  "Adversarial training (AT) <a href='#page-9-1'>[23]</a>, which aims to minimize\n",
-  "the model's risk under the worst-case perturbations, is cur-\n",
-  "rently the most effective approach for improving the robust-\n",
-  "ness of deep neural networks. For a given neural network\n",
-  "f(x, w) with parameters w, the optimization objective of\n",
-  "AT can be formulated as follows:\n"
- ]
-}
-```
-
-Output:
-analysis: The inline math in the lines is not in LaTeX format and is not surrounded by <math>...</math> tags.
-```json
-{
- "corrected_lines": [
-  "Adversarial training (AT) <a href='#page-9-1'>[23]</a>, which aims to minimize\n",
-  "the model's risk under the worst-case perturbations, is cur-\n",
-  "rently the most effective approach for improving the robust-\n",
-  "ness of deep neural networks. For a given neural network\n",
-  "<math>f(x, w)</math> with parameters <math>w</math>, the optimization objective of\n",
-  "AT can be formulated as follows:\n"
- ]
-}
-```
-
-**Input:**
-```json
-{extracted_lines}
-```
-"""
-
-    def inference_blocks(self, document: Document) -> List[List[BlockData]]:
-        blocks = []
-        for page in document.pages:
-            for block in page.contained_blocks(document, self.block_types):
-                if block.formats and "math" in block.formats:
-                    blocks.append({
-                        "page": page,
-                        "block": block
-                    })
-
-
-        out_blocks = []
-        for i in range(0, len(blocks), self.math_line_batch_size):
-            batch = blocks[i:i + self.math_line_batch_size]
-            out_blocks.append(batch)
-        return out_blocks
-
-    def get_block_lines(self, block: Block, document: Document) -> Tuple[list, list]:
-        text_lines = block.contained_blocks(document, (BlockTypes.Line,))
-        extracted_lines = [line.formatted_text(document) for line in text_lines]
-        return text_lines, extracted_lines
-
-    def combine_images(self, images: List[Image.Image]):
-        widths, heights = zip(*(i.size for i in images))
-        total_width = max(widths)
-        total_height = sum(heights) + 5 * len(images)
-
-        new_im = Image.new('RGB', (total_width, total_height), (255, 255, 255))
-
-        y_offset = 0
-        for im in images:
-            new_im.paste(im, (0, y_offset))
-            y_offset += im.size[1] + 5
-
-        return new_im
-
-    def block_prompts(self, document: Document) -> List[PromptData]:
-        prompt_data = []
-        for block_data in self.inference_blocks(document):
-            blocks: List[Line] = [b["block"] for b in block_data]
-            pages = [b["page"] for b in block_data]
-            block_lines = [block.formatted_text(document) for block in blocks]
-
-            prompt = (
-                self.text_math_rewriting_prompt
-                  .replace("{extracted_lines}",json.dumps({"extracted_lines": block_lines}, indent=2))
-                  .replace("{line_count}", str(len(block_lines)))
-            )
-            images = [self.extract_image(document, block, remove_blocks=self.image_remove_blocks) for block in blocks]
-            image = self.combine_images(images)
-
-            prompt_data.append({
-                "prompt": prompt,
-                "image": image,
-                "block": blocks[0],
-                "schema": LLMTextSchema,
-                "page": pages[0],
-                "additional_data": {"blocks": blocks, "pages": pages}
-            })
-        return prompt_data
-
-
-    def rewrite_block(self, response: dict, prompt_data: PromptData, document: Document):
-        blocks = prompt_data["additional_data"]["blocks"]
-        pages = prompt_data["additional_data"]["pages"]
-
-        if not response or "corrected_lines" not in response:
-            blocks[0].update_metadata(llm_error_count=1)
-            return
-
-        corrected_lines = response["corrected_lines"]
-        balanced_math = all([line.count("<math") == line.count("</math>") for line in corrected_lines])
-        if any([
-            not corrected_lines,
-            len(corrected_lines) != len(blocks),
-            not balanced_math
-        ]):
-            blocks[0].update_metadata(llm_error_count=1)
-            return
-
-        for text_line, page, corrected_text in zip(blocks, pages, corrected_lines):
-            text_line.structure = []
-            add_math_spans_to_line(corrected_text, text_line, page)
-
-class LLMTextSchema(BaseModel):
-    analysis: str
-    corrected_lines: List[str]
\ No newline at end of file
diff --git a/marker/scripts/streamlit_app.py b/marker/scripts/streamlit_app.py
index 2445e12b..928e94b5 100644
--- a/marker/scripts/streamlit_app.py
+++ b/marker/scripts/streamlit_app.py
@@ -256,3 +256,5 @@ def block_display(image: Image, blocks: dict | None = None, dpi=96):
             layout_image_path = os.path.join(debug_data_path, f"layout_page_{first_page}.png")
             img = Image.open(layout_image_path)
             st.image(img, caption="Layout debug image", use_container_width=True)
+        st.write("Raw output:")
+        st.code(text, language=output_format)
diff --git a/poetry.lock b/poetry.lock
index 9045cf29..c2c3d54c 100644
--- a/poetry.lock
+++ b/poetry.lock
@@ -3115,30 +3115,6 @@ files = [
     {file = "nvidia_nvtx_cu12-12.4.127-py3-none-win_amd64.whl", hash = "sha256:641dccaaa1139f3ffb0d3164b4b84f9d253397e38246a4f2f36728b48566d485"},
 ]
 
-[[package]]
-name = "opencv-python-headless"
-version = "4.11.0.86"
-description = "Wrapper package for OpenCV python bindings."
-optional = false
-python-versions = ">=3.6"
-files = [
-    {file = "opencv-python-headless-4.11.0.86.tar.gz", hash = "sha256:996eb282ca4b43ec6a3972414de0e2331f5d9cda2b41091a49739c19fb843798"},
-    {file = "opencv_python_headless-4.11.0.86-cp37-abi3-macosx_13_0_arm64.whl", hash = "sha256:48128188ade4a7e517237c8e1e11a9cdf5c282761473383e77beb875bb1e61ca"},
-    {file = "opencv_python_headless-4.11.0.86-cp37-abi3-macosx_13_0_x86_64.whl", hash = "sha256:a66c1b286a9de872c343ee7c3553b084244299714ebb50fbdcd76f07ebbe6c81"},
-    {file = "opencv_python_headless-4.11.0.86-cp37-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:6efabcaa9df731f29e5ea9051776715b1bdd1845d7c9530065c7951d2a2899eb"},
-    {file = "opencv_python_headless-4.11.0.86-cp37-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:0e0a27c19dd1f40ddff94976cfe43066fbbe9dfbb2ec1907d66c19caef42a57b"},
-    {file = "opencv_python_headless-4.11.0.86-cp37-abi3-win32.whl", hash = "sha256:f447d8acbb0b6f2808da71fddd29c1cdd448d2bc98f72d9bb78a7a898fc9621b"},
-    {file = "opencv_python_headless-4.11.0.86-cp37-abi3-win_amd64.whl", hash = "sha256:6c304df9caa7a6a5710b91709dd4786bf20a74d57672b3c31f7033cc638174ca"},
-]
-
-[package.dependencies]
-numpy = [
-    {version = ">=1.26.0", markers = "python_version >= \"3.12\""},
-    {version = ">=1.23.5", markers = "python_version >= \"3.11\" and python_version < \"3.12\""},
-    {version = ">=1.21.4", markers = "python_version >= \"3.10\" and platform_system == \"Darwin\" and python_version < \"3.11\""},
-    {version = ">=1.21.2", markers = "platform_system != \"Darwin\" and python_version >= \"3.10\" and python_version < \"3.11\""},
-]
-
 [[package]]
 name = "openpyxl"
 version = "3.1.5"
@@ -3228,9 +3204,9 @@ files = [
 
 [package.dependencies]
 numpy = [
-    {version = ">=1.26.0", markers = "python_version >= \"3.12\""},
-    {version = ">=1.23.2", markers = "python_version == \"3.11\""},
     {version = ">=1.22.4", markers = "python_version < \"3.11\""},
+    {version = ">=1.23.2", markers = "python_version == \"3.11\""},
+    {version = ">=1.26.0", markers = "python_version >= \"3.12\""},
 ]
 python-dateutil = ">=2.8.2"
 pytz = ">=2020.1"
@@ -5076,27 +5052,16 @@ snowflake = ["snowflake-connector-python (>=3.3.0)", "snowflake-snowpark-python[
 
 [[package]]
 name = "surya-ocr"
-version = "0.12.1"
+version = "0.13.0"
 description = "OCR, layout, reading order, and table recognition in 90+ languages"
 optional = false
-python-versions = "<4.0,>=3.10"
-files = [
-    {file = "surya_ocr-0.12.1-py3-none-any.whl", hash = "sha256:703362d808994576e7cd297c731c74af552c1be831cbb933968736a187d89ac3"},
-    {file = "surya_ocr-0.12.1.tar.gz", hash = "sha256:4b37d94db3747f843c23c9da4f146c436de514945297bdebfc394971c03f6340"},
-]
+python-versions = "*"
+files = []
+develop = false
 
-[package.dependencies]
-click = ">=8.1.8,<9.0.0"
-filetype = ">=1.2.0,<2.0.0"
-opencv-python-headless = ">=4.11.0.86,<5.0.0.0"
-pillow = ">=10.2.0,<11.0.0"
-platformdirs = ">=4.3.6,<5.0.0"
-pydantic = ">=2.5.3,<3.0.0"
-pydantic-settings = ">=2.1.0,<3.0.0"
-pypdfium2 = "4.30.0"
-python-dotenv = ">=1.0.0,<2.0.0"
-torch = ">=2.5.1,<3.0.0"
-transformers = ">=4.41.0,<5.0.0"
+[package.source]
+type = "directory"
+url = "../surya"
 
 [[package]]
 name = "sympy"
@@ -6112,4 +6077,4 @@ full = ["ebooklib", "mammoth", "openpyxl", "python-pptx", "weasyprint"]
 [metadata]
 lock-version = "2.0"
 python-versions = "^3.10"
-content-hash = "7d16a2f72bebb9a040c49a446dc9951b9e50bc88a007e2f646d0786206cdfde4"
+content-hash = "e03ee53a4be2afc3661be448bc16a102b47273a57125b37e54184f88030b233b"
diff --git a/pyproject.toml b/pyproject.toml
index 03481ab0..31fdc570 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -26,7 +26,7 @@ torch = "^2.5.1"
 tqdm = "^4.66.1"
 ftfy = "^6.1.1"
 rapidfuzz = "^3.8.1"
-surya-ocr = "~0.12.1"
+surya-ocr = "~0.13.0"
 regex = "^2024.4.28"
 pdftext = "~0.6.1"
 markdownify = "^0.13.1"
diff --git a/tests/processors/test_inline_math.py b/tests/processors/test_inline_math.py
index a609ba38..7145f1c9 100644
--- a/tests/processors/test_inline_math.py
+++ b/tests/processors/test_inline_math.py
@@ -3,7 +3,7 @@
 import pytest
 
 from marker.processors.llm.llm_meta import LLMSimpleBlockMetaProcessor
-from marker.processors.llm.llm_text import LLMTextProcessor
+from marker.processors.llm.llm_inlinemath import LLMInlineMathLinesProcessor
 from marker.schema import BlockTypes
 
 
@@ -20,7 +20,7 @@ def test_llm_text_processor(pdf_document, mocker):
     mock_cls.return_value = {"corrected_lines": corrected_lines}
 
     config = {"use_llm": True, "gemini_api_key": "test"}
-    processor_lst = [LLMTextProcessor(config)]
+    processor_lst = [LLMInlineMathLinesProcessor(config)]
     processor = LLMSimpleBlockMetaProcessor(processor_lst, mock_cls, config)
     processor(pdf_document)
 
diff --git a/tests/processors/test_llm_processors.py b/tests/processors/test_llm_processors.py
index ecdff9db..d87828ff 100644
--- a/tests/processors/test_llm_processors.py
+++ b/tests/processors/test_llm_processors.py
@@ -8,7 +8,7 @@
 from marker.processors.llm.llm_image_description import LLMImageDescriptionProcessor
 from marker.processors.llm.llm_meta import LLMSimpleBlockMetaProcessor
 from marker.processors.llm.llm_table import LLMTableProcessor
-from marker.processors.llm.llm_text import LLMTextProcessor
+from marker.processors.llm.llm_inlinemath import LLMInlineMathLinesProcessor
 from marker.processors.table import TableProcessor
 from marker.renderers.markdown import MarkdownRenderer
 from marker.schema import BlockTypes

From aad174b7af9379729d59ea4a80aa37f2120a6a68 Mon Sep 17 00:00:00 2001
From: Vik Paruchuri <vik.paruchuri@gmail.com>
Date: Fri, 28 Feb 2025 15:48:28 -0500
Subject: [PATCH 44/46] Skip some lines

---
 marker/processors/llm/llm_inlinemath.py | 12 +++++++++++-
 1 file changed, 11 insertions(+), 1 deletion(-)

diff --git a/marker/processors/llm/llm_inlinemath.py b/marker/processors/llm/llm_inlinemath.py
index 8c32adc4..054ef653 100644
--- a/marker/processors/llm/llm_inlinemath.py
+++ b/marker/processors/llm/llm_inlinemath.py
@@ -82,8 +82,18 @@ class LLMInlineMathLinesProcessor(BaseLLMSimpleBlockProcessor):
     def inference_blocks(self, document: Document) -> List[List[BlockData]]:
         blocks = []
         for page in document.pages:
+            page_children = [p for p in page.children if p.structure]
             for block in page.contained_blocks(document, self.block_types):
-                if block.formats and "math" in block.formats:
+                # Ensure the line isn't an orphan, and that the parent hasn't already been inferenced (assigned html)
+                has_parent = any([
+                    (
+                        block.id in parent.structure
+                        and not getattr(parent, "html", None)
+                    )
+                    for parent in page_children
+                ])
+
+                if block.formats and "math" in block.formats and has_parent:
                     blocks.append({
                         "page": page,
                         "block": block

From 4b2c1a494332ac2f3782db7cf049d1a415c359fc Mon Sep 17 00:00:00 2001
From: Vik Paruchuri <vik.paruchuri@gmail.com>
Date: Fri, 28 Feb 2025 17:12:38 -0500
Subject: [PATCH 45/46] Update surya

---
 marker/processors/llm/llm_table.py | 11 ++++---
 poetry.lock                        | 51 +++++++++++++++++++++++++-----
 2 files changed, 49 insertions(+), 13 deletions(-)

diff --git a/marker/processors/llm/llm_table.py b/marker/processors/llm/llm_table.py
index 264ae29b..68049536 100644
--- a/marker/processors/llm/llm_table.py
+++ b/marker/processors/llm/llm_table.py
@@ -39,13 +39,14 @@ class LLMTableProcessor(BaseLLMComplexBlockProcessor):
         "Default is a string containing the Gemini rewriting prompt."
     ] = """You are a text correction expert specializing in accurately reproducing text from images.
 You will receive an image and an html representation of the table in the image.
-Your task is to correct any errors in the html representation.  The html representation should be as faithful to the original table as possible.  The table may be rotated, but ensure the html representation is not rotated.  Make sure to include HTML for the full table, including the opening and closing table tags.
+Your task is to correct any errors in the html representation.  The html representation should be as faithful to the original table image as possible.  The table image may be rotated, but ensure the html representation is not rotated.  Make sure to include HTML for the full table, including the opening and closing table tags.
 
 Some guidelines:
-- Make sure to reproduce the original values as faithfully as possible.
+- Reproduce the original values from the image as faithfully as possible.  
+- There may be stray characters in the html representation that don't match the image - fix these.
 - Ensure column headers match the correct column values.
-- If you see any math in a table cell, fence it with the <math> tag.  Block math should be fenced with <math display="block">.
-- Replace any images with a description, like "Image: [description]".
+- If you see any inline math in a table cell, fence it with the <math> tag.  Block math should be fenced with <math display="block">.
+- Replace any images in table cells with a description, like "Image: [description]".
 - Only use the tags th, td, tr, br, span, sup, sub, i, b, math, and table.  Only use the attributes display, style, colspan, and rowspan if necessary.  You can use br to break up text lines in cells.
 - Make sure the columns and rows match the image faithfully, and are easily readable and interpretable by a human.
 
@@ -71,8 +72,8 @@ class LLMTableProcessor(BaseLLMComplexBlockProcessor):
 </table>
 ```
 Output:
+comparison: The image shows a table with 2 rows and 3 columns.  The text and formatting of the html table matches the image.  The column headers match the correct column values.
 ```html
-Comparison: The image shows a table with 2 rows and 3 columns.  The text and formatting of the html table matches the image.  The column headers match the correct column values.
 No corrections needed.
 ```
 **Input:**
diff --git a/poetry.lock b/poetry.lock
index c2c3d54c..46a15094 100644
--- a/poetry.lock
+++ b/poetry.lock
@@ -3115,6 +3115,30 @@ files = [
     {file = "nvidia_nvtx_cu12-12.4.127-py3-none-win_amd64.whl", hash = "sha256:641dccaaa1139f3ffb0d3164b4b84f9d253397e38246a4f2f36728b48566d485"},
 ]
 
+[[package]]
+name = "opencv-python-headless"
+version = "4.11.0.86"
+description = "Wrapper package for OpenCV python bindings."
+optional = false
+python-versions = ">=3.6"
+files = [
+    {file = "opencv-python-headless-4.11.0.86.tar.gz", hash = "sha256:996eb282ca4b43ec6a3972414de0e2331f5d9cda2b41091a49739c19fb843798"},
+    {file = "opencv_python_headless-4.11.0.86-cp37-abi3-macosx_13_0_arm64.whl", hash = "sha256:48128188ade4a7e517237c8e1e11a9cdf5c282761473383e77beb875bb1e61ca"},
+    {file = "opencv_python_headless-4.11.0.86-cp37-abi3-macosx_13_0_x86_64.whl", hash = "sha256:a66c1b286a9de872c343ee7c3553b084244299714ebb50fbdcd76f07ebbe6c81"},
+    {file = "opencv_python_headless-4.11.0.86-cp37-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:6efabcaa9df731f29e5ea9051776715b1bdd1845d7c9530065c7951d2a2899eb"},
+    {file = "opencv_python_headless-4.11.0.86-cp37-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:0e0a27c19dd1f40ddff94976cfe43066fbbe9dfbb2ec1907d66c19caef42a57b"},
+    {file = "opencv_python_headless-4.11.0.86-cp37-abi3-win32.whl", hash = "sha256:f447d8acbb0b6f2808da71fddd29c1cdd448d2bc98f72d9bb78a7a898fc9621b"},
+    {file = "opencv_python_headless-4.11.0.86-cp37-abi3-win_amd64.whl", hash = "sha256:6c304df9caa7a6a5710b91709dd4786bf20a74d57672b3c31f7033cc638174ca"},
+]
+
+[package.dependencies]
+numpy = [
+    {version = ">=1.26.0", markers = "python_version >= \"3.12\""},
+    {version = ">=1.23.5", markers = "python_version >= \"3.11\" and python_version < \"3.12\""},
+    {version = ">=1.21.4", markers = "python_version >= \"3.10\" and platform_system == \"Darwin\" and python_version < \"3.11\""},
+    {version = ">=1.21.2", markers = "platform_system != \"Darwin\" and python_version >= \"3.10\" and python_version < \"3.11\""},
+]
+
 [[package]]
 name = "openpyxl"
 version = "3.1.5"
@@ -3204,9 +3228,9 @@ files = [
 
 [package.dependencies]
 numpy = [
-    {version = ">=1.22.4", markers = "python_version < \"3.11\""},
-    {version = ">=1.23.2", markers = "python_version == \"3.11\""},
     {version = ">=1.26.0", markers = "python_version >= \"3.12\""},
+    {version = ">=1.23.2", markers = "python_version == \"3.11\""},
+    {version = ">=1.22.4", markers = "python_version < \"3.11\""},
 ]
 python-dateutil = ">=2.8.2"
 pytz = ">=2020.1"
@@ -5055,13 +5079,24 @@ name = "surya-ocr"
 version = "0.13.0"
 description = "OCR, layout, reading order, and table recognition in 90+ languages"
 optional = false
-python-versions = "*"
-files = []
-develop = false
+python-versions = "<4.0,>=3.10"
+files = [
+    {file = "surya_ocr-0.13.0-py3-none-any.whl", hash = "sha256:b22124a6a0d6e3547b43d1a818c086efc55a7a84e4ce3f7561f793306ad44534"},
+    {file = "surya_ocr-0.13.0.tar.gz", hash = "sha256:93b866ced75d9599dbbb0f746498f114d10f5063f588013a27e5192b4d307cc9"},
+]
 
-[package.source]
-type = "directory"
-url = "../surya"
+[package.dependencies]
+click = ">=8.1.8,<9.0.0"
+filetype = ">=1.2.0,<2.0.0"
+opencv-python-headless = ">=4.11.0.86,<5.0.0.0"
+pillow = ">=10.2.0,<11.0.0"
+platformdirs = ">=4.3.6,<5.0.0"
+pydantic = ">=2.5.3,<3.0.0"
+pydantic-settings = ">=2.1.0,<3.0.0"
+pypdfium2 = "4.30.0"
+python-dotenv = ">=1.0.0,<2.0.0"
+torch = ">=2.5.1,<3.0.0"
+transformers = ">=4.41.0,<5.0.0"
 
 [[package]]
 name = "sympy"

From b586f78fc6fba09cab4dec2def25639c98e29775 Mon Sep 17 00:00:00 2001
From: Vik Paruchuri <vik.paruchuri@gmail.com>
Date: Fri, 28 Feb 2025 18:20:08 -0500
Subject: [PATCH 46/46] Bump pdftext, fix tests

---
 README.md                               | 2 +-
 marker/processors/llm/llm_equation.py   | 6 +++++-
 marker/providers/document.py            | 3 +--
 marker/providers/epub.py                | 6 +++---
 marker/providers/powerpoint.py          | 7 ++++---
 marker/providers/spreadsheet.py         | 8 +++-----
 poetry.lock                             | 8 ++++----
 pyproject.toml                          | 2 +-
 tests/builders/test_garbled_pdf.py      | 2 +-
 tests/builders/test_ocr_pipeline.py     | 2 +-
 tests/conftest.py                       | 2 +-
 tests/processors/test_llm_processors.py | 2 +-
 12 files changed, 26 insertions(+), 24 deletions(-)

diff --git a/README.md b/README.md
index 5e7b806e..0ad95be9 100644
--- a/README.md
+++ b/README.md
@@ -66,7 +66,7 @@ Install with:
 pip install marker-pdf
 ```
 
-By default, marker will work on PDFs and images.  If you also want to use marker on XLSX, DOCX, HTML, etc, you will need to run:
+If you want to use marker on documents other than PDFs, you will need to install additional dependencies with:
 
 ```shell
 pip install marker-pdf[full]
diff --git a/marker/processors/llm/llm_equation.py b/marker/processors/llm/llm_equation.py
index 73a62256..3ed8273c 100644
--- a/marker/processors/llm/llm_equation.py
+++ b/marker/processors/llm/llm_equation.py
@@ -74,8 +74,12 @@ def inference_blocks(self, document: Document) -> List[BlockData]:
         for block_data in blocks:
             block = block_data["block"]
             page = block_data["page"]
+
             # If we redo inline math, we redo all equations
-            if block.polygon.height / page.polygon.height < self.min_equation_height and not self.redo_inline_math:
+            if all([
+                block.polygon.height / page.polygon.height < self.min_equation_height,
+                not self.redo_inline_math
+            ]):
                 continue
             out_blocks.append(block_data)
         return out_blocks
diff --git a/marker/providers/document.py b/marker/providers/document.py
index 00798061..4dfd1038 100644
--- a/marker/providers/document.py
+++ b/marker/providers/document.py
@@ -1,11 +1,9 @@
 import base64
-import logging
 import os
 import re
 import tempfile
 from io import BytesIO
 
-import mammoth
 from PIL import Image
 
 from marker.providers.pdf import PdfProvider
@@ -69,6 +67,7 @@ def __del__(self):
 
     def convert_docx_to_pdf(self, filepath: str):
         from weasyprint import CSS, HTML
+        import mammoth
 
         with open(filepath, "rb") as docx_file:
             # we convert the docx to HTML
diff --git a/marker/providers/epub.py b/marker/providers/epub.py
index 3cb25110..4dd75ed0 100644
--- a/marker/providers/epub.py
+++ b/marker/providers/epub.py
@@ -2,9 +2,7 @@
 import os
 import tempfile
 
-import ebooklib
 from bs4 import BeautifulSoup
-from ebooklib import epub
 
 from marker.providers.pdf import PdfProvider
 
@@ -67,6 +65,8 @@ def __del__(self):
 
     def convert_epub_to_pdf(self, filepath):
         from weasyprint import CSS, HTML
+        from ebooklib import epub
+        import ebooklib
 
         ebook = epub.read_epub(filepath)
 
@@ -104,7 +104,7 @@ def convert_epub_to_pdf(self, filepath):
         full_style = ''.join([css])  # + styles)
 
         # we convert the epub to HTML
-        result = HTML(string=html_content, base_url=filepath).write_pdf(
+        HTML(string=html_content, base_url=filepath).write_pdf(
             self.temp_pdf_path,
             stylesheets=[CSS(string=full_style), self.get_font_css()]
         )
diff --git a/marker/providers/powerpoint.py b/marker/providers/powerpoint.py
index 2bc6dcee..4c7e6987 100644
--- a/marker/providers/powerpoint.py
+++ b/marker/providers/powerpoint.py
@@ -3,9 +3,6 @@
 import tempfile
 import traceback
 
-from pptx import Presentation
-from pptx.enum.shapes import MSO_SHAPE_TYPE, PP_PLACEHOLDER
-
 from marker.providers.pdf import PdfProvider
 
 css = '''
@@ -63,6 +60,8 @@ def __del__(self):
 
     def convert_pptx_to_pdf(self, filepath):
         from weasyprint import CSS, HTML
+        from pptx import Presentation
+        from pptx.enum.shapes import MSO_SHAPE_TYPE
 
         pptx = Presentation(filepath)
 
@@ -112,6 +111,7 @@ def _handle_group(self, group_shape) -> str:
         """
         Recursively handle shapes in a group. Returns HTML string for the entire group.
         """
+        from pptx.enum.shapes import MSO_SHAPE_TYPE
 
         group_parts = []
         for shape in group_shape.shapes:
@@ -140,6 +140,7 @@ def _handle_text(self, shape) -> str:
         Processes shape text, including bullet/numbered list detection and placeholders
         (title, subtitle, etc.). Returns HTML for the text block(s).
         """
+        from pptx.enum.shapes import PP_PLACEHOLDER
 
         # Distinguish placeholders to see if it's a title or subtitle
         label_html_tag = "p"
diff --git a/marker/providers/spreadsheet.py b/marker/providers/spreadsheet.py
index c544f7d9..ddad461d 100644
--- a/marker/providers/spreadsheet.py
+++ b/marker/providers/spreadsheet.py
@@ -1,9 +1,6 @@
 import os
 import tempfile
 
-from openpyxl import load_workbook
-from openpyxl.worksheet.worksheet import Worksheet
-
 from marker.providers.pdf import PdfProvider
 
 css = '''
@@ -52,6 +49,7 @@ def __del__(self):
 
     def convert_xlsx_to_pdf(self, filepath: str):
         from weasyprint import CSS, HTML
+        from openpyxl import load_workbook
 
         html = ""
         workbook = load_workbook(filepath)
@@ -69,7 +67,7 @@ def convert_xlsx_to_pdf(self, filepath: str):
         )
 
     @staticmethod
-    def _get_merged_cell_ranges(sheet: Worksheet):
+    def _get_merged_cell_ranges(sheet):
         merged_info = {}
         for merged_range in sheet.merged_cells.ranges:
             min_col, min_row, max_col, max_row = merged_range.bounds
@@ -80,7 +78,7 @@ def _get_merged_cell_ranges(sheet: Worksheet):
             }
         return merged_info
 
-    def _excel_to_html_table(self, sheet: Worksheet):
+    def _excel_to_html_table(self, sheet):
         merged_cells = self._get_merged_cell_ranges(sheet)
 
         html = f'<table>'
diff --git a/poetry.lock b/poetry.lock
index 46a15094..95d01d76 100644
--- a/poetry.lock
+++ b/poetry.lock
@@ -3289,13 +3289,13 @@ testing = ["docopt", "pytest"]
 
 [[package]]
 name = "pdftext"
-version = "0.6.1"
+version = "0.6.2"
 description = "Extract structured text from pdfs quickly"
 optional = false
 python-versions = "<4.0,>=3.10"
 files = [
-    {file = "pdftext-0.6.1-py3-none-any.whl", hash = "sha256:9c437a05262277dede2f6953eebc7b46d7393bb11ee373267814af2aa5e02e4d"},
-    {file = "pdftext-0.6.1.tar.gz", hash = "sha256:ffec41064804e157b48b76c834051ed7a5aa456257b78d9b87a5e8f54cebe307"},
+    {file = "pdftext-0.6.2-py3-none-any.whl", hash = "sha256:905d11e62d548e307933c25865a69c8e993947bb5b40b1535b0a2aa8f07a71d4"},
+    {file = "pdftext-0.6.2.tar.gz", hash = "sha256:ff5b92462ac03ae63a23429384ae123d45c162dcda30e7bf2c5c92a6b208c9de"},
 ]
 
 [package.dependencies]
@@ -6112,4 +6112,4 @@ full = ["ebooklib", "mammoth", "openpyxl", "python-pptx", "weasyprint"]
 [metadata]
 lock-version = "2.0"
 python-versions = "^3.10"
-content-hash = "e03ee53a4be2afc3661be448bc16a102b47273a57125b37e54184f88030b233b"
+content-hash = "4609798f8e0c4bc0c7a9ab4bcb6f92289ea03bdd902a1c35c12699f874f67298"
diff --git a/pyproject.toml b/pyproject.toml
index 31fdc570..d2342860 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -28,7 +28,7 @@ ftfy = "^6.1.1"
 rapidfuzz = "^3.8.1"
 surya-ocr = "~0.13.0"
 regex = "^2024.4.28"
-pdftext = "~0.6.1"
+pdftext = "~0.6.2"
 markdownify = "^0.13.1"
 click = "^8.1.7"
 markdown2 = "^2.5.2"
diff --git a/tests/builders/test_garbled_pdf.py b/tests/builders/test_garbled_pdf.py
index 281ec84a..4fa4a49f 100644
--- a/tests/builders/test_garbled_pdf.py
+++ b/tests/builders/test_garbled_pdf.py
@@ -15,7 +15,7 @@ def test_garbled_pdf(pdf_document, detection_model, recognition_model, table_rec
 
     table_cell = pdf_document.pages[0].get_block(table_block.structure[0])
     assert table_cell.block_type == BlockTypes.Line
-    assert table_cell.structure[0] == "/page/0/Span/2"
+    assert table_cell.structure[0] == "/page/0/Span/3"
 
     # We don't OCR in the initial pass, only with the TableProcessor
     processor = TableProcessor(detection_model, recognition_model, table_rec_model)
diff --git a/tests/builders/test_ocr_pipeline.py b/tests/builders/test_ocr_pipeline.py
index f079517e..4a43dc54 100644
--- a/tests/builders/test_ocr_pipeline.py
+++ b/tests/builders/test_ocr_pipeline.py
@@ -23,7 +23,7 @@ def _ocr_pipeline_test(pdf_document):
     # Makes sure the OCR bbox is being scaled to the same scale as the layout boxes
     text_lines = first_page.contained_blocks(pdf_document, (BlockTypes.Line,))
     text_blocks = first_page.contained_blocks(pdf_document, (BlockTypes.Text, BlockTypes.TextInlineMath))
-    assert len(text_lines) == 71
+    assert len(text_lines) == 84
 
     # Ensure the bbox sizes match up
     max_line_position = max([line.polygon.y_end for line in text_lines])
diff --git a/tests/conftest.py b/tests/conftest.py
index a40f8a21..e4c083c7 100644
--- a/tests/conftest.py
+++ b/tests/conftest.py
@@ -147,7 +147,7 @@ def llm_service(request, config):
 def temp_image():
     img = Image.new("RGB", (512, 512), color="white")
     draw = ImageDraw.Draw(img)
-    draw.text((10, 10), "Hello, World!", fill="black")
+    draw.text((10, 10), "Hello, World!", fill="black", font_size=24)
     with tempfile.NamedTemporaryFile(suffix=".png") as f:
         img.save(f.name)
         f.flush()
diff --git a/tests/processors/test_llm_processors.py b/tests/processors/test_llm_processors.py
index d87828ff..b85b6321 100644
--- a/tests/processors/test_llm_processors.py
+++ b/tests/processors/test_llm_processors.py
@@ -168,7 +168,7 @@ def test_llm_complex_region_processor(pdf_document):
 def test_multi_llm_processors(pdf_document):
     description = "<math>This is an image description.  And here is a lot of writing about it.</math>" * 10
     mock_cls = Mock()
-    mock_cls.return_value = {"image_description": description, "html_equation": description}
+    mock_cls.return_value = {"image_description": description, "corrected_equation": description}
 
     config = {"use_llm": True, "gemini_api_key": "test", "extract_images": False, "min_equation_height": .001}
     processor_lst = [LLMImageDescriptionProcessor(config), LLMEquationProcessor(config)]