Merge branch 'alpha' into 'feat/support-native-gemini-embedding'

relay/channel/gemini/adaptor.go

@@ -119,6 +119,7 @@ func (a *Adaptor) GetRequestURL(info *relaycommon.RelayInfo) (string, error) {
 			action = "batchEmbedContents"
 		}
 		return fmt.Sprintf("%s/%s/models/%s:%s", info.BaseUrl, version, info.UpstreamModelName, action), nil
+		return fmt.Sprintf("%s/%s/models/%s:batchEmbedContents", info.BaseUrl, version, info.UpstreamModelName), nil
 	}
 
 	action := "generateContent"
@@ -163,29 +164,35 @@ func (a *Adaptor) ConvertEmbeddingRequest(c *gin.Context, info *relaycommon.Rela
 	if len(inputs) == 0 {
 		return nil, errors.New("input is empty")
 	}
-
-	// only process the first input
-	geminiRequest := dto.GeminiEmbeddingRequest{
-		Content: dto.GeminiChatContent{
-			Parts: []dto.GeminiPart{
-				{
-					Text: inputs[0],
+	// process all inputs
+	geminiRequests := make([]map[string]interface{}, 0, len(inputs))
+	for _, input := range inputs {
+		geminiRequest := map[string]interface{}{
+			"model": fmt.Sprintf("models/%s", info.UpstreamModelName),
+			"content": dto.GeminiChatContent{
+				Parts: []dto.GeminiPart{
+					{
+						Text: input,
+					},
 				},
 			},
-		},
-	}
-
-	// set specific parameters for different models
-	// https://ai.google.dev/api/embeddings?hl=zh-cn#method:-models.embedcontent
-	switch info.UpstreamModelName {
-	case "text-embedding-004":
-		// except embedding-001 supports setting `OutputDimensionality`
-		if request.Dimensions > 0 {
-			geminiRequest.OutputDimensionality = request.Dimensions
 		}
+
+		// set specific parameters for different models
+		// https://ai.google.dev/api/embeddings?hl=zh-cn#method:-models.embedcontent
+		switch info.UpstreamModelName {
+		case "text-embedding-004", "gemini-embedding-exp-03-07", "gemini-embedding-001":
+			// Only newer models introduced after 2024 support OutputDimensionality
+			if request.Dimensions > 0 {
+				geminiRequest["outputDimensionality"] = request.Dimensions
+			}
+		}
+		geminiRequests = append(geminiRequests, geminiRequest)
 	}
 
-	return geminiRequest, nil
+	return map[string]interface{}{
+		"requests": geminiRequests,
+	}, nil
 }
 
 func (a *Adaptor) ConvertOpenAIResponsesRequest(c *gin.Context, info *relaycommon.RelayInfo, request dto.OpenAIResponsesRequest) (any, error) {

relay/channel/gemini/relay-gemini.go

@@ -1071,7 +1071,7 @@ func GeminiEmbeddingHandler(c *gin.Context, info *relaycommon.RelayInfo, resp *h
 		return nil, types.NewOpenAIError(readErr, types.ErrorCodeBadResponseBody, http.StatusInternalServerError)
 	}
 
-	var geminiResponse dto.GeminiEmbeddingResponse
+	var geminiResponse dto.GeminiBatchEmbeddingResponse
 	if jsonErr := common.Unmarshal(responseBody, &geminiResponse); jsonErr != nil {
 		return nil, types.NewOpenAIError(jsonErr, types.ErrorCodeBadResponseBody, http.StatusInternalServerError)
 	}
@@ -1079,14 +1079,16 @@ func GeminiEmbeddingHandler(c *gin.Context, info *relaycommon.RelayInfo, resp *h
 	// convert to openai format response
 	openAIResponse := dto.OpenAIEmbeddingResponse{
 		Object: "list",
-		Data: []dto.OpenAIEmbeddingResponseItem{
-			{
-				Object:    "embedding",
-				Embedding: geminiResponse.Embedding.Values,
-				Index:     0,
-			},
-		},
-		Model: info.UpstreamModelName,
+		Data:   make([]dto.OpenAIEmbeddingResponseItem, 0, len(geminiResponse.Embeddings)),
+		Model:  info.UpstreamModelName,
+	}
+
+	for i, embedding := range geminiResponse.Embeddings {
+		openAIResponse.Data = append(openAIResponse.Data, dto.OpenAIEmbeddingResponseItem{
+			Object:    "embedding",
+			Embedding: embedding.Values,
+			Index:     i,
+		})
 	}
 
 	// calculate usage