feat: Support transforming reasoning_content returned by Qwen to OpenAI contract (#1791)

plugins/wasm-go/extensions/ai-proxy/provider/model.go

@@ -1,6 +1,9 @@
 package provider
 
-import "strings"
+import (
+	"fmt"
+	"strings"
+)
 
 const (
 	streamEventIdItemKey        = "id:"
@@ -110,9 +113,16 @@ type chatCompletionChoice struct {
 }
 
 type usage struct {
-	PromptTokens     int `json:"prompt_tokens,omitempty"`
-	CompletionTokens int `json:"completion_tokens,omitempty"`
-	TotalTokens      int `json:"total_tokens,omitempty"`
+	PromptTokens            int                      `json:"prompt_tokens,omitempty"`
+	CompletionTokens        int                      `json:"completion_tokens,omitempty"`
+	TotalTokens             int                      `json:"total_tokens,omitempty"`
+	CompletionTokensDetails *completionTokensDetails `json:"completion_tokens_details,omitempty"`
+}
+
+type completionTokensDetails struct {
+	ReasoningTokens          int `json:"reasoning_tokens,omitempty"`
+	AcceptedPredictionTokens int `json:"accepted_prediction_tokens,omitempty"`
+	RejectedPredictionTokens int `json:"rejected_prediction_tokens,omitempty"`
 }
 
 type chatMessage struct {
@@ -126,6 +136,24 @@ type chatMessage struct {
 	Refusal          string                 `json:"refusal,omitempty"`
 }
 
+func (m *chatMessage) handleReasoningContent(reasoningContentMode string) {
+	if m.ReasoningContent == "" {
+		return
+	}
+	switch reasoningContentMode {
+	case reasoningBehaviorIgnore:
+		m.ReasoningContent = ""
+		break
+	case reasoningBehaviorConcat:
+		m.Content = fmt.Sprintf("%v\n%v", m.ReasoningContent, m.Content)
+		m.ReasoningContent = ""
+		break
+	case reasoningBehaviorPassThrough:
+	default:
+		break
+	}
+}
+
 type messageContent struct {
 	Type     string    `json:"type,omitempty"`
 	Text     string    `json:"text"`
@@ -138,6 +166,9 @@ type imageUrl struct {
 }
 
 func (m *chatMessage) IsEmpty() bool {
+	if m.ReasoningContent != "" {
+		return false
+	}
 	if m.IsStringContent() && m.Content != "" {
 		return false
 	}

plugins/wasm-go/extensions/ai-proxy/provider/provider.go

@@ -85,6 +85,10 @@ const (
 	objectChatCompletion      = "chat.completion"
 	objectChatCompletionChunk = "chat.completion.chunk"
 
+	reasoningBehaviorPassThrough = "passthrough"
+	reasoningBehaviorIgnore      = "ignore"
+	reasoningBehaviorConcat      = "concat"
+
 	wildcard = "*"
 
 	defaultTimeout = 2 * 60 * 1000 // ms
@@ -190,6 +194,9 @@ type ProviderConfig struct {
 	// @Title zh-CN 失败请求重试
 	// @Description zh-CN 对失败的请求立即进行重试
 	retryOnFailure *retryOnFailure `required:"false" yaml:"retryOnFailure" json:"retryOnFailure"`
+	// @Title zh-CN 推理内容处理方式
+	// @Description zh-CN 如何处理大模型服务返回的推理内容。目前支持以下取值：passthrough（正常输出推理内容）、ignore（不输出推理内容）、concat（将推理内容拼接在常规输出内容之前）。默认为 normal。仅支持通义千问服务。
+	reasoningContentMode string `required:"false" yaml:"reasoningContentMode" json:"reasoningContentMode"`
 	// @Title zh-CN 基于OpenAI协议的自定义后端URL
 	// @Description zh-CN 仅适用于支持 openai 协议的服务。
 	openaiCustomUrl string `required:"false" yaml:"openaiCustomUrl" json:"openaiCustomUrl"`
@@ -359,6 +366,20 @@ func (c *ProviderConfig) FromJson(json gjson.Result) {
 		}
 	}
 
+	c.reasoningContentMode = json.Get("reasoningContentMode").String()
+	if c.reasoningContentMode == "" {
+		c.reasoningContentMode = reasoningBehaviorPassThrough
+	} else {
+		c.reasoningContentMode = strings.ToLower(c.reasoningContentMode)
+		switch c.reasoningContentMode {
+		case reasoningBehaviorPassThrough, reasoningBehaviorIgnore, reasoningBehaviorConcat:
+			break
+		default:
+			c.reasoningContentMode = reasoningBehaviorPassThrough
+			break
+		}
+	}
+
 	failoverJson := json.Get("failover")
 	c.failover = &failover{
 		enabled: false,

plugins/wasm-go/extensions/ai-proxy/provider/qwen.go

@@ -357,7 +357,7 @@ func (m *qwenProvider) buildQwenTextGenerationRequest(ctx wrapper.HttpContext, o
 func (m *qwenProvider) buildChatCompletionResponse(ctx wrapper.HttpContext, qwenResponse *qwenTextGenResponse) *chatCompletionResponse {
 	choices := make([]chatCompletionChoice, 0, len(qwenResponse.Output.Choices))
 	for _, qwenChoice := range qwenResponse.Output.Choices {
-		message := qwenMessageToChatMessage(qwenChoice.Message)
+		message := qwenMessageToChatMessage(qwenChoice.Message, m.config.reasoningContentMode)
 		choices = append(choices, chatCompletionChoice{
 			Message:      &message,
 			FinishReason: qwenChoice.FinishReason,
@@ -395,7 +395,8 @@ func (m *qwenProvider) buildChatCompletionStreamingResponse(ctx wrapper.HttpCont
 	finished := qwenChoice.FinishReason != "" && qwenChoice.FinishReason != "null"
 	message := qwenChoice.Message
 
-	deltaContentMessage := &chatMessage{Role: message.Role, Content: message.Content}
+	deltaContentMessage := &chatMessage{Role: message.Role, Content: message.Content, ReasoningContent: message.ReasoningContent}
+	deltaContentMessage.handleReasoningContent(m.config.reasoningContentMode)
 	deltaToolCallsMessage := &chatMessage{Role: message.Role, ToolCalls: append([]toolCall{}, message.ToolCalls...)}
 	if !incrementalStreaming {
 		for _, tc := range message.ToolCalls {
@@ -430,6 +431,11 @@ func (m *qwenProvider) buildChatCompletionStreamingResponse(ctx wrapper.HttpCont
 					}
 				}
 			}
+			if message.ReasoningContent == "" {
+				message.ReasoningContent = pushedMessage.ReasoningContent
+			} else {
+				deltaContentMessage.ReasoningContent = util.StripPrefix(deltaContentMessage.ReasoningContent, pushedMessage.ReasoningContent)
+			}
 			if len(deltaToolCallsMessage.ToolCalls) > 0 && pushedMessage.ToolCalls != nil {
 				for i, tc := range deltaToolCallsMessage.ToolCalls {
 					if i >= len(pushedMessage.ToolCalls) {
@@ -690,13 +696,16 @@ type qwenTextEmbeddings struct {
 	Embedding []float64 `json:"embedding"`
 }
 
-func qwenMessageToChatMessage(qwenMessage qwenMessage) chatMessage {
-	return chatMessage{
-		Name:      qwenMessage.Name,
-		Role:      qwenMessage.Role,
-		Content:   qwenMessage.Content,
-		ToolCalls: qwenMessage.ToolCalls,
+func qwenMessageToChatMessage(qwenMessage qwenMessage, reasoningContentMode string) chatMessage {
+	msg := chatMessage{
+		Name:             qwenMessage.Name,
+		Role:             qwenMessage.Role,
+		Content:          qwenMessage.Content,
+		ReasoningContent: qwenMessage.ReasoningContent,
+		ToolCalls:        qwenMessage.ToolCalls,
 	}
+	msg.handleReasoningContent(reasoningContentMode)
+	return msg
 }
 
 func (m *qwenMessage) IsStringContent() bool {