jiazhizhong/higress
1
0
Fork 0
mirror of https://github.com/alibaba/higress.git synced 2026-05-24 04:37:25 +08:00
Code Issues Packages Projects Releases Wiki Activity

Vertex structured outputs (#3649)

Browse Source
This commit is contained in:
woody
2026-03-30 17:26:55 +08:00
committed by GitHub
parent 889ea67013
commit cd8ed99db5
4 changed files with 720 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files

167
plugins/wasm-go/extensions/ai-proxy/provider/vertex.go
View File

@@ -259,12 +259,12 @@ func (v *vertexProvider) OnRequestBody(ctx wrapper.HttpContext, apiName ApiName,
if v.isOpenAICompatibleMode() {
ctx.SetContext(contextOpenAICompatibleMarker, true)
body, err := v.onOpenAICompatibleRequestBody(ctx, apiName, body, headers)
headers.Set("Content-Length", fmt.Sprint(len(body)))
util.ReplaceRequestHeaders(headers)
_ = proxywasm.ReplaceHttpRequestBody(body)
if err != nil {
return types.ActionContinue, err
}
headers.Set("Content-Length", fmt.Sprint(len(body)))
util.ReplaceRequestHeaders(headers)
_ = proxywasm.ReplaceHttpRequestBody(body)
// OpenAI 兼容模式需要 OAuth token
cached, err := v.getToken()
if cached {
@@ -277,6 +277,9 @@ func (v *vertexProvider) OnRequestBody(ctx wrapper.HttpContext, apiName ApiName,
}
body, err := v.TransformRequestBodyHeaders(ctx, apiName, body, headers)
if err != nil {
return types.ActionContinue, err
}
headers.Set("Content-Length", fmt.Sprint(len(body)))
if v.isExpressMode() {
@@ -284,15 +287,12 @@ func (v *vertexProvider) OnRequestBody(ctx wrapper.HttpContext, apiName ApiName,
headers.Del("Authorization")
util.ReplaceRequestHeaders(headers)
_ = proxywasm.ReplaceHttpRequestBody(body)
return types.ActionContinue, err
return types.ActionContinue, nil
}
// 标准模式: 需要获取 OAuth token
util.ReplaceRequestHeaders(headers)
_ = proxywasm.ReplaceHttpRequestBody(body)
if err != nil {
return types.ActionContinue, err
}
cached, err := v.getToken()
if cached {
return types.ActionContinue, nil
@@ -369,7 +369,10 @@ func (v *vertexProvider) onChatCompletionRequestBody(ctx wrapper.HttpContext, bo
path := v.getRequestPath(ApiNameChatCompletion, request.Model, request.Stream)
util.OverwriteRequestPathHeader(headers, path)
vertexRequest := v.buildVertexChatRequest(request)
vertexRequest, err := v.buildVertexChatRequest(request)
if err != nil {
return nil, err
}
return json.Marshal(vertexRequest)
}
}
@@ -971,7 +974,7 @@ func (v *vertexProvider) getOpenAICompatibleRequestPath() string {
return fmt.Sprintf(vertexOpenAICompatiblePathTemplate, v.config.vertexProjectId, v.config.vertexRegion)
}
func (v *vertexProvider) buildVertexChatRequest(request *chatCompletionRequest) *vertexChatRequest {
func (v *vertexProvider) buildVertexChatRequest(request *chatCompletionRequest) (*vertexChatRequest, error) {
safetySettings := make([]vertexChatSafetySetting, 0)
for category, threshold := range v.config.geminiSafetySetting {
safetySettings = append(safetySettings, vertexChatSafetySetting{
@@ -1006,6 +1009,9 @@ func (v *vertexProvider) buildVertexChatRequest(request *chatCompletionRequest)
}
vertexRequest.GenerationConfig.ThinkingConfig = thinkingConfig
}
if err := v.applyResponseFormatToGenerationConfig(request.ResponseFormat, &vertexRequest.GenerationConfig, request.Model); err != nil {
return nil, err
}
if request.Tools != nil {
functions := make([]function, 0, len(request.Tools))
for _, tool := range request.Tools {
@@ -1091,7 +1097,130 @@ func (v *vertexProvider) buildVertexChatRequest(request *chatCompletionRequest)
}
}
return &vertexRequest
return &vertexRequest, nil
}
// applyResponseFormatToGenerationConfig maps OpenAI response_format into Vertex generationConfig.
// The mapping is strict for type=json_schema to avoid silently breaking structured-output contracts.
func (v *vertexProvider) applyResponseFormatToGenerationConfig(responseFormat map[string]interface{}, generationConfig *vertexChatGenerationConfig, model string) error {
if generationConfig == nil || len(responseFormat) == 0 {
return nil
}
// NOTE: Gemini 2.0 structured output requires propertyOrdering.
// Because gemini-2.0-* is legacy and rarely used, we intentionally do not implement
// propertyOrdering synthesis here; instead we ignore response_format and keep request
// as non-structured output for stability and minimal conversion behavior.
if requiresPropertyOrderingForModel(model) {
return nil
}
responseFormatType, _ := responseFormat["type"].(string)
responseFormatType = strings.ToLower(responseFormatType)
switch responseFormatType {
case "":
// Be tolerant for non-standard clients that pass schema directly in response_format.
if isJSONSchemaMap(responseFormat) {
generationConfig.ResponseMimeType = util.MimeTypeApplicationJson
generationConfig.ResponseSchema = responseFormat
}
case "json_object":
generationConfig.ResponseMimeType = util.MimeTypeApplicationJson
case "json_schema":
schema := extractOpenAIJSONSchema(responseFormat)
if len(schema) == 0 {
return fmt.Errorf("invalid response_format.json_schema: missing schema object")
}
generationConfig.ResponseMimeType = util.MimeTypeApplicationJson
generationConfig.ResponseSchema = schema
case "text":
// Vertex defaults to text output when no response mime/schema is provided.
default:
// Be tolerant for non-standard usage where response_format itself is a JSON schema.
if isJSONSchemaType(responseFormatType) && isJSONSchemaMap(responseFormat) {
generationConfig.ResponseMimeType = util.MimeTypeApplicationJson
generationConfig.ResponseSchema = responseFormat
}
}
return nil
}
func extractOpenAIJSONSchema(responseFormat map[string]interface{}) map[string]interface{} {
jsonSchemaValue, ok := responseFormat["json_schema"]
if !ok {
return nil
}
jsonSchemaMap, ok := jsonSchemaValue.(map[string]interface{})
if !ok {
return nil
}
// OpenAI canonical format:
// {
// "type":"json_schema",
// "json_schema":{"name":"...","strict":true,"schema":{...}}
// }
if nestedSchemaValue, ok := jsonSchemaMap["schema"]; ok {
if nestedSchema, ok := nestedSchemaValue.(map[string]interface{}); ok {
return nestedSchema
}
}
// Tolerate non-standard format where json_schema itself is the schema.
if isJSONSchemaMap(jsonSchemaMap) {
return jsonSchemaMap
}
return nil
}
func isJSONSchemaType(value string) bool {
switch strings.ToLower(value) {
case "object", "array", "string", "number", "integer", "boolean", "null":
return true
default:
return false
}
}
func isJSONSchemaMap(schema map[string]interface{}) bool {
if len(schema) == 0 {
return false
}
if typeValue, ok := schema["type"].(string); ok && isJSONSchemaType(typeValue) {
return true
}
// Schema might omit "type" and still be valid for specific cases.
schemaKeys := []string{
"anyOf",
"enum",
"format",
"items",
"maximum",
"maxItems",
"minimum",
"minItems",
"nullable",
"properties",
"description",
"propertyOrdering",
"required",
}
for _, key := range schemaKeys {
if _, ok := schema[key]; ok {
return true
}
}
return false
}
func requiresPropertyOrderingForModel(model string) bool {
model = strings.ToLower(model)
return strings.HasPrefix(model, "gemini-2.0-")
}
func (v *vertexProvider) buildEmbeddingRequest(request *embeddingsRequest) *vertexEmbeddingRequest {
@@ -1170,14 +1299,16 @@ type vertexChatSafetySetting struct {
}
type vertexChatGenerationConfig struct {
Temperature float64 `json:"temperature,omitempty"`
TopP float64 `json:"topP,omitempty"`
TopK int `json:"topK,omitempty"`
CandidateCount int `json:"candidateCount,omitempty"`
MaxOutputTokens int `json:"maxOutputTokens,omitempty"`
ThinkingConfig vertexThinkingConfig `json:"thinkingConfig,omitempty"`
ResponseModalities []string `json:"responseModalities,omitempty"`
ImageConfig *vertexImageConfig `json:"imageConfig,omitempty"`
Temperature float64 `json:"temperature,omitempty"`
TopP float64 `json:"topP,omitempty"`
TopK int `json:"topK,omitempty"`
CandidateCount int `json:"candidateCount,omitempty"`
MaxOutputTokens int `json:"maxOutputTokens,omitempty"`
ThinkingConfig vertexThinkingConfig `json:"thinkingConfig,omitempty"`
ResponseMimeType string `json:"responseMimeType,omitempty"`
ResponseSchema map[string]interface{} `json:"responseSchema,omitempty"`
ResponseModalities []string `json:"responseModalities,omitempty"`
ImageConfig *vertexImageConfig `json:"imageConfig,omitempty"`
}
type vertexImageConfig struct {