Fins SSA GenAI Offering

The goal of this repo is to develop and deliver GenAI solutions to enable and accelerate customer GenAI PoC Projects with Databricks

Table of Content

• Requirements
• PoC Accelerator Short Assets
  • Data Ingestion and Preprocessing Architecture Patterns
  • End to End GenAI Application Architecture Patterns
• When to Use
  • Getting Started
• Resources
• Limitations

Requirements

• Databricks workspace with Serverless and Unity Catalog enabled
• Python 3.9+

PoC Accelerator Short Assets

Data Ingestion and Preprocessing Architecture Patterns

Input Data Types	Input Data Store	chunking performed	OSS Technolgoy	Component Asset
JSON Text Transcripts	Unity Catalog Volum	None	None	Json Data Ingestion with DLT (python), DLT (SQL)
PDF Doc (with tables)	Unity Catalog Volum	Unstructured chunking strategy	Unstructured	PDF Doc Ingestion
Image Doc (text extraction)	Unity Catalog Volum	Unstructured chunking strategy	Unstructured	Image Doc Ingestion

End to End GenAI Application Architecture Patterns

Input Data	Model	Tasks	GenAI Use Case	Orchestration	Customer Persona	PoC Template
JSON Text Transcripts	Foundation LLM (e.g. Llama3p1)	ummarization, Sentiment, classification	AI Function, DBSQL Agent	DLT, LangChain	Data Analyist, Data Scientist	Call Center Transcript Analytics with AI
JSON Text Transcripts	Foundation LLM (e.g. Llama3p1)	ummarization, Sentiment	RAG	DLT, LangChain	Data Scientist, MLE, Data Engineer	Call Center Transcript RAG Apps
wav Audio	Foundation LLM (e.g. Llama3p1)	Speech Transcription, Summarization, Sentiment	RAG	DLT, LangChain	Data Scientist, MLE, Data Engineer	Call Center Audio to Text RAG Apps
PDF Documents	Foundation LLM (e.g. Llama3p1)	Unstructured Data Processing, Name Entity Recognition	NER	DLT, Function calling	Data Scientist, MLE, Data Engineer	PDF_Doc Ingestion

When to Use

You have a business use case that can potentially apply generative AI technology and fall into one of the PoC accelerator template. You have access to a unity catalog enabled Databricks Workspace.

You may have some existing data available in the workspace to use as input data. If you don't have any data, the PoC accelerator templates contains synthetic sample datasets to enable the demonstration of genAI application's functionalities

Getting Started

Clone this repo and add the repo to your Databricks Workspace. Refer to Databricks Repo Setup for instuctions on how to create Databricks repo on your workspace

1. Got into the folder of the selected PoC accelerator template
2. Review the architecture diagram in the README
3. Start with the instruction notebook
4. Follow the instructions in the instruction notebook.
5. Most of notebook can run by click Run/Run ALL but some may require additional steps of using databricks UI so be sure to read the instruction

Resources

• Databricks GenAI Cookbook
• Databricks Foundation Model
• Model Serving
• Vector Search
• Inference Table
• Delta Live Table
• Databricks Python SDK
• MLFlow

Limitations

• The PoC accelerator template is designed for use Unit Catalog managed workspace only.
• The synthetic dataset provided by Databricks are generated algorithmatically based on assumptions and they are not real-life data.
• Delta Live Table technology from Databricks is used in some of PoC Accelerator Template, Currently the live table (a.k.a materialized view) from Delta Live Table cannot only be accessed by shared clusters, therefore, a copy of the materialized views are being used in some of notebooks. The limitation will be addressed in the future product releases