update

Signed-off-by: yxia216 <[email protected]>

Author: hustxiayang

internal/apischema/gcp/gcp.go

@@ -36,3 +36,13 @@ type GenerateContentRequest struct {
 	// https://github.com/googleapis/go-genai/blob/6a8184fcaf8bf15f0c566616a7b356560309be9b/types.go#L1057
 	SafetySettings []*genai.SafetySetting `json:"safetySettings,omitempty"`
 }
+
+type EmbedContentRequest struct {
+	// Contains the multipart content of a message.
+	//
+	// https://github.com/googleapis/go-genai/blob/6a8184fcaf8bf15f0c566616a7b356560309be9b/types.go#L858
+	Contents []genai.Content `json:"contents"`
+	// Tool details of a tool that the model may use to generate a response.
+
+	Config *genai.EmbedContentConfig `json:"config,omitempty"`
+}

internal/apischema/openai/openai.go

@@ -1525,6 +1525,11 @@ type EmbeddingCompletionRequest struct {
 	User *string `json:"user,omitempty"`
 }
 
+// GetModel implements ModelName interface
+func (e *EmbeddingCompletionRequest) GetModel() string {
+	return e.Model
+}
+
 // EmbeddingChatRequest represents a request structure for embeddings API. This is not a standard openai, but just extend the request to have messages/chat like completion requests
 type EmbeddingChatRequest struct {
 	// Messages: A list of messages comprising the conversation so far.
@@ -1550,10 +1555,28 @@ type EmbeddingChatRequest struct {
 	User *string `json:"user,omitempty"`
 }
 
-type EmbedddingRequest interface {
+// GetModel implements ModelProvider interface
+func (e *EmbeddingChatRequest) GetModel() string {
+	return e.Model
+}
+
+type EmbeddingRequest interface {
 	EmbeddingCompletionRequest | EmbeddingChatRequest
 }
 
+// ModelName interface for types that can provide a model name
+type ModelName interface {
+	GetModel() string
+}
+
+// GetModelFromEmbeddingRequest extracts the model name from any EmbeddingRequest type
+func GetModelFromEmbeddingRequest[T EmbeddingRequest](req *T) string {
+	if mp, ok := any(*req).(ModelName); ok {
+		return mp.GetModel()
+	}
+	return ""
+}
+
 // EmbeddingResponse represents a response from /v1/embeddings.
 // https://platform.openai.com/docs/api-reference/embeddings/object
 type EmbeddingResponse struct {

internal/translator/openai_azureopenai_embeddings.go

@@ -33,7 +33,7 @@ type openAIToAzureOpenAITranslatorV1Embedding struct {
 }
 
 // RequestBody implements [OpenAIEmbeddingTranslator.RequestBody].
-func (o *openAIToAzureOpenAITranslatorV1Embedding) RequestBody(original []byte, req *openai.EmbeddingRequest, onRetry bool) (
+func (o *openAIToAzureOpenAITranslatorV1Embedding) RequestBody(original []byte, req *openai.EmbeddingCompletionRequest, onRetry bool) (
 	newHeaders []internalapi.Header, newBody []byte, err error,
 ) {
 	modelName := req.Model

internal/translator/openai_embeddings.go

@@ -36,7 +36,7 @@ type openAIToOpenAITranslatorV1Embedding struct {
 }
 
 // RequestBody implements [OpenAIEmbeddingTranslator.RequestBody].
-func (o *openAIToOpenAITranslatorV1Embedding) RequestBody(original []byte, _ *openai.EmbeddingRequest, onRetry bool) (
+func (o *openAIToOpenAITranslatorV1Embedding) RequestBody(original []byte, _ *openai.EmbeddingCompletionRequest, onRetry bool) (
 	newHeaders []internalapi.Header, newBody []byte, err error,
 ) {
 	if o.modelNameOverride != "" {

internal/translator/openai_gcpvertexai_embeddings.go

@@ -0,0 +1,115 @@
+// Copyright Envoy AI Gateway Authors
+// SPDX-License-Identifier: Apache-2.0
+// The full text of the Apache license is available in the LICENSE file at
+// the root of the repo.
+
+package translator
+
+import (
+	"encoding/json"
+	"fmt"
+	"strconv"
+
+	"github.com/envoyproxy/ai-gateway/internal/apischema/gcp"
+	"github.com/envoyproxy/ai-gateway/internal/apischema/openai"
+	"github.com/envoyproxy/ai-gateway/internal/internalapi"
+)
+
+// NewEmbeddingOpenAIToAzureOpenAITranslator implements [Factory] for OpenAI to Azure OpenAI translation
+// for embeddings.
+func NewEmbeddingOpenAIToGCPVertexAITranslator(requestModel internalapi.RequestModel, modelNameOverride internalapi.ModelNameOverride) OpenAIEmbeddingTranslator {
+	return &openAIToGCPVertexAITranslatorV1Embedding{
+		apiVersion: apiVersion,
+		openAIToOpenAITranslatorV1Embedding: openAIToOpenAITranslatorV1Embedding{
+			modelNameOverride: modelNameOverride,
+		},
+	}
+}
+
+// openAIToGCPVertexAITranslatorV1Embedding implements [OpenAIEmbeddingTranslator] for /embeddings.
+type openAIToGCPVertexAITranslatorV1Embedding[T openai.EmbeddingRequest] struct {
+	requestModel internalapi.RequestModel
+	openAIToOpenAITranslatorV1Embedding
+}
+
+
+
+func InputToGeminiConent(input openai.EmbeddingRequestInput){
+	 switch v := input.Value.(type) {
+      case string:
+
+          return v, "string", nil
+      case []string:
+          // Array of text inputs
+          return v, "string_array", nil
+      case []int64:
+          // Array of token IDs
+          return v, "token_array", nil
+      case [][]int64:
+          // Array of token ID arrays
+          return v, "token_array_batch", nil
+      default:
+          return nil, "unknown", fmt.Errorf("unsupported input type: %T", v)
+      }
+
+
+}
+
+// openAIToGCPVertexAITranslatorV1Embedding converts an OpenAI EmbeddingRequest to a GCP Gemini GenerateContentRequest.
+func openAIEmbeddingCompletionToGeminiMessage(openAIReq *openai.EmbeddingCompletionRequest, requestModel internalapi.RequestModel) (*gcp.EmbedContentRequest, error) {
+	// Convert OpenAI EmbeddingRequest's input to Gemini Contents
+	contents, err := InputToGeminiConent(openAIReq.Input, requestModel)
+	if err != nil {
+		return nil, err
+	}
+
+	// Convert generation config.
+	embedConfig,, err := openAIReqToGeminiGenerationConfig(openAIReq, requestModel)
+	if err != nil {
+		return nil, fmt.Errorf("error converting generation config: %w", err)
+	}
+
+	gcr := gcp.EmbedContentRequest{
+		Contents:          contents,
+		Config:  embedConfig,
+	}
+
+	return &gcr, nil
+}
+
+// RequestBody implements [OpenAIEmbeddingTranslator.RequestBody].
+func (o *openAIToGCPVertexAITranslatorV1Embedding[T]) RequestBody(original []byte, req *T, onRetry bool) (
+	newHeaders []internalapi.Header, newBody []byte, err error,
+) {
+
+	o.requestModel = openai.GetModelFromEmbeddingRequest(req)
+	if o.modelNameOverride != "" {
+		// Use modelName override if set.
+		o.requestModel = o.modelNameOverride
+	}
+
+	// Choose the correct endpoint based on streaming.
+	var path string
+
+	path = buildGCPModelPathSuffix(gcpModelPublisherGoogle, o.requestModel, gcpMethodGenerateContent)
+
+	switch any(*req).(type) {
+	case openai.EmbeddingCompletionRequest:
+		gcpReq, err := openAIEmbeddingCompletionToGeminiMessage(openAIReq, o.requestModel)
+	case openai.EmbeddingChatRequest:
+		gcpReq, err := openAIEmbeddingChatToGeminiMessage(openAIReq, o.requestModel)
+
+	default:
+		return nil, nil, fmt.Errorf("request body is wrong: %w", err)
+	}
+
+	newBody, err = json.Marshal(gcpReq)
+	if err != nil {
+		return nil, nil, fmt.Errorf("error marshaling Gemini request: %w", err)
+	}
+	newHeaders = []internalapi.Header{
+		{pathHeaderName, path},
+		{contentLengthHeaderName, strconv.Itoa(len(newBody))},
+	}
+	return
+}

internal/translator/translator.go

@@ -78,7 +78,7 @@ type (
 	// OpenAIChatCompletionTranslator translates the OpenAI's /chat/completions endpoint.
 	OpenAIChatCompletionTranslator = Translator[openai.ChatCompletionRequest, tracing.ChatCompletionSpan]
 	// OpenAIEmbeddingTranslator translates the OpenAI's /embeddings endpoint.
-	OpenAIEmbeddingTranslator = Translator[openai.EmbeddingRequest, tracing.EmbeddingsSpan]
+	OpenAIEmbeddingTranslator = Translator[openai.EmbeddingCompletionRequest, tracing.EmbeddingsSpan]
 	// OpenAICompletionTranslator translates the OpenAI's /completions endpoint.
 	OpenAICompletionTranslator = Translator[openai.CompletionRequest, tracing.CompletionSpan]
 	// CohereRerankTranslator translates the Cohere's /v2/rerank endpoint.