Merge branch 'sashabaranov:master' into master

README.md

@@ -7,7 +7,7 @@ This library provides unofficial Go clients for [OpenAI API](https://platform.op
 
 * ChatGPT
 * GPT-3, GPT-4
-* DALL·E 2
+* DALL·E 2, DALL·E 3
 * Whisper
 
 ## Installation

audio.go

@@ -27,8 +27,14 @@ const (
 	AudioResponseFormatVTT         AudioResponseFormat = "vtt"
 )
 
+type TranscriptionTimestampGranularity string
+
+const (
+	TranscriptionTimestampGranularityWord    TranscriptionTimestampGranularity = "word"
+	TranscriptionTimestampGranularitySegment TranscriptionTimestampGranularity = "segment"
+)
+
 // AudioRequest represents a request structure for audio API.
-// ResponseFormat is not supported for now. We only return JSON text, which may be sufficient.
 type AudioRequest struct {
 	Model string
 
@@ -38,10 +44,11 @@ type AudioRequest struct {
 	// Reader is an optional io.Reader when you do not want to use an existing file.
 	Reader io.Reader
 
-	Prompt      string // For translation, it should be in English
-	Temperature float32
-	Language    string // For translation, just do not use it. It seems "en" works, not confirmed...
-	Format      AudioResponseFormat
+	Prompt                 string
+	Temperature            float32
+	Language               string // Only for transcription.
+	Format                 AudioResponseFormat
+	TimestampGranularities []TranscriptionTimestampGranularity // Only for transcription.
 }
 
 // AudioResponse represents a response structure for audio API.
@@ -62,6 +69,11 @@ type AudioResponse struct {
 		NoSpeechProb     float64 `json:"no_speech_prob"`
 		Transient        bool    `json:"transient"`
 	} `json:"segments"`
+	Words []struct {
+		Word  string  `json:"word"`
+		Start float64 `json:"start"`
+		End   float64 `json:"end"`
+	} `json:"words"`
 	Text string `json:"text"`
 
 	httpHeader
@@ -179,6 +191,15 @@ func audioMultipartForm(request AudioRequest, b utils.FormBuilder) error {
 		}
 	}
 
+	if len(request.TimestampGranularities) > 0 {
+		for _, tg := range request.TimestampGranularities {
+			err = b.WriteField("timestamp_granularities[]", string(tg))
+			if err != nil {
+				return fmt.Errorf("writing timestamp_granularities[]: %w", err)
+			}
+		}
+	}
+
 	// Close the multipart writer
 	return b.Close()
 }

audio_api_test.go

@@ -105,6 +105,10 @@ func TestAudioWithOptionalArgs(t *testing.T) {
 				Temperature: 0.5,
 				Language:    "zh",
 				Format:      openai.AudioResponseFormatSRT,
+				TimestampGranularities: []openai.TranscriptionTimestampGranularity{
+					openai.TranscriptionTimestampGranularitySegment,
+					openai.TranscriptionTimestampGranularityWord,
+				},
 			}
 			_, err := tc.createFn(ctx, req)
 			checks.NoError(t, err, "audio API error")

audio_test.go

@@ -24,6 +24,10 @@ func TestAudioWithFailingFormBuilder(t *testing.T) {
 		Temperature: 0.5,
 		Language:    "en",
 		Format:      AudioResponseFormatSRT,
+		TimestampGranularities: []TranscriptionTimestampGranularity{
+			TranscriptionTimestampGranularitySegment,
+			TranscriptionTimestampGranularityWord,
+		},
 	}
 
 	mockFailedErr := fmt.Errorf("mock form builder fail")
@@ -47,7 +51,7 @@ func TestAudioWithFailingFormBuilder(t *testing.T) {
 		return nil
 	}
 
-	failOn := []string{"model", "prompt", "temperature", "language", "response_format"}
+	failOn := []string{"model", "prompt", "temperature", "language", "response_format", "timestamp_granularities[]"}
 	for _, failingField := range failOn {
 		failForField = failingField
 		mockFailedErr = fmt.Errorf("mock form builder fail on field %s", failingField)

chat.go

@@ -216,6 +216,16 @@ type ChatCompletionRequest struct {
 	Tools        []Tool `json:"tools,omitempty"`
 	// This can be either a string or an ToolChoice object.
 	ToolChoice any `json:"tool_choice,omitempty"`
+	// Options for streaming response. Only set this when you set stream: true.
+	StreamOptions *StreamOptions `json:"stream_options,omitempty"`
+}
+
+type StreamOptions struct {
+	// If set, an additional chunk will be streamed before the data: [DONE] message.
+	// The usage field on this chunk shows the token usage statistics for the entire request,
+	// and the choices field will always be an empty array.
+	// All other chunks will also include a usage field, but with a null value.
+	IncludeUsage bool `json:"include_usage,omitempty"`
 }
 
 type ToolType string

chat_stream.go

@@ -33,6 +33,10 @@ type ChatCompletionStreamResponse struct {
 	SystemFingerprint   string                       `json:"system_fingerprint"`
 	PromptAnnotations   []PromptAnnotation           `json:"prompt_annotations,omitempty"`
 	PromptFilterResults []PromptFilterResult         `json:"prompt_filter_results,omitempty"`
+	// An optional field that will only be present when you set stream_options: {"include_usage": true} in your request.
+	// When present, it contains a null value except for the last chunk which contains the token usage statistics
+	// for the entire request.
+	Usage *Usage `json:"usage,omitempty"`
 }
 
 // ChatCompletionStream

chat_stream_test.go

@@ -388,6 +388,120 @@ func TestAzureCreateChatCompletionStreamRateLimitError(t *testing.T) {
 	}
 }
 
+func TestCreateChatCompletionStreamStreamOptions(t *testing.T) {
+	client, server, teardown := setupOpenAITestServer()
+	defer teardown()
+
+	server.RegisterHandler("/v1/chat/completions", func(w http.ResponseWriter, _ *http.Request) {
+		w.Header().Set("Content-Type", "text/event-stream")
+
+		// Send test responses
+		var dataBytes []byte
+		//nolint:lll
+		data := `{"id":"1","object":"completion","created":1598069254,"model":"gpt-3.5-turbo","system_fingerprint": "fp_d9767fc5b9","choices":[{"index":0,"delta":{"content":"response1"},"finish_reason":"max_tokens"}],"usage":null}`
+		dataBytes = append(dataBytes, []byte("data: "+data+"\n\n")...)
+
+		//nolint:lll
+		data = `{"id":"2","object":"completion","created":1598069255,"model":"gpt-3.5-turbo","system_fingerprint": "fp_d9767fc5b9","choices":[{"index":0,"delta":{"content":"response2"},"finish_reason":"max_tokens"}],"usage":null}`
+		dataBytes = append(dataBytes, []byte("data: "+data+"\n\n")...)
+
+		//nolint:lll
+		data = `{"id":"3","object":"completion","created":1598069256,"model":"gpt-3.5-turbo","system_fingerprint": "fp_d9767fc5b9","choices":[],"usage":{"prompt_tokens":1,"completion_tokens":1,"total_tokens":2}}`
+		dataBytes = append(dataBytes, []byte("data: "+data+"\n\n")...)
+
+		dataBytes = append(dataBytes, []byte("data: [DONE]\n\n")...)
+
+		_, err := w.Write(dataBytes)
+		checks.NoError(t, err, "Write error")
+	})
+
+	stream, err := client.CreateChatCompletionStream(context.Background(), openai.ChatCompletionRequest{
+		MaxTokens: 5,
+		Model:     openai.GPT3Dot5Turbo,
+		Messages: []openai.ChatCompletionMessage{
+			{
+				Role:    openai.ChatMessageRoleUser,
+				Content: "Hello!",
+			},
+		},
+		Stream: true,
+		StreamOptions: &openai.StreamOptions{
+			IncludeUsage: true,
+		},
+	})
+	checks.NoError(t, err, "CreateCompletionStream returned error")
+	defer stream.Close()
+
+	expectedResponses := []openai.ChatCompletionStreamResponse{
+		{
+			ID:                "1",
+			Object:            "completion",
+			Created:           1598069254,
+			Model:             openai.GPT3Dot5Turbo,
+			SystemFingerprint: "fp_d9767fc5b9",
+			Choices: []openai.ChatCompletionStreamChoice{
+				{
+					Delta: openai.ChatCompletionStreamChoiceDelta{
+						Content: "response1",
+					},
+					FinishReason: "max_tokens",
+				},
+			},
+		},
+		{
+			ID:                "2",
+			Object:            "completion",
+			Created:           1598069255,
+			Model:             openai.GPT3Dot5Turbo,
+			SystemFingerprint: "fp_d9767fc5b9",
+			Choices: []openai.ChatCompletionStreamChoice{
+				{
+					Delta: openai.ChatCompletionStreamChoiceDelta{
+						Content: "response2",
+					},
+					FinishReason: "max_tokens",
+				},
+			},
+		},
+		{
+			ID:                "3",
+			Object:            "completion",
+			Created:           1598069256,
+			Model:             openai.GPT3Dot5Turbo,
+			SystemFingerprint: "fp_d9767fc5b9",
+			Choices:           []openai.ChatCompletionStreamChoice{},
+			Usage: &openai.Usage{
+				PromptTokens:     1,
+				CompletionTokens: 1,
+				TotalTokens:      2,
+			},
+		},
+	}
+
+	for ix, expectedResponse := range expectedResponses {
+		b, _ := json.Marshal(expectedResponse)
+		t.Logf("%d: %s", ix, string(b))
+
+		receivedResponse, streamErr := stream.Recv()
+		checks.NoError(t, streamErr, "stream.Recv() failed")
+		if !compareChatResponses(expectedResponse, receivedResponse) {
+			t.Errorf("Stream response %v is %v, expected %v", ix, receivedResponse, expectedResponse)
+		}
+	}
+
+	_, streamErr := stream.Recv()
+	if !errors.Is(streamErr, io.EOF) {
+		t.Errorf("stream.Recv() did not return EOF in the end: %v", streamErr)
+	}
+
+	_, streamErr = stream.Recv()
+
+	checks.ErrorIs(t, streamErr, io.EOF, "stream.Recv() did not return EOF when the stream is finished")
+	if !errors.Is(streamErr, io.EOF) {
+		t.Errorf("stream.Recv() did not return EOF when the stream is finished: %v", streamErr)
+	}
+}
+
 // Helper funcs.
 func compareChatResponses(r1, r2 openai.ChatCompletionStreamResponse) bool {
 	if r1.ID != r2.ID || r1.Object != r2.Object || r1.Created != r2.Created || r1.Model != r2.Model {
@@ -401,6 +515,15 @@ func compareChatResponses(r1, r2 openai.ChatCompletionStreamResponse) bool {
 			return false
 		}
 	}
+	if r1.Usage != nil || r2.Usage != nil {
+		if r1.Usage == nil || r2.Usage == nil {
+			return false
+		}
+		if r1.Usage.PromptTokens != r2.Usage.PromptTokens || r1.Usage.CompletionTokens != r2.Usage.CompletionTokens ||
+			r1.Usage.TotalTokens != r2.Usage.TotalTokens {
+			return false
+		}
+	}
 	return true
 }
 

completion.go

@@ -22,6 +22,8 @@ const (
 	GPT432K               = "gpt-4-32k"
 	GPT40613              = "gpt-4-0613"
 	GPT40314              = "gpt-4-0314"
+	GPT4o                 = "gpt-4o"
+	GPT4o20240513         = "gpt-4o-2024-05-13"
 	GPT4Turbo             = "gpt-4-turbo"
 	GPT4Turbo20240409     = "gpt-4-turbo-2024-04-09"
 	GPT4Turbo0125         = "gpt-4-0125-preview"
@@ -82,6 +84,8 @@ var disabledModelsForEndpoints = map[string]map[string]bool{
 		GPT3Dot5Turbo16K:     true,
 		GPT3Dot5Turbo16K0613: true,
 		GPT4:                 true,
+		GPT4o:                true,
+		GPT4o20240513:        true,
 		GPT4TurboPreview:     true,
 		GPT4VisionPreview:    true,
 		GPT4Turbo1106:        true,