fix: #352

app/client/platforms/openai.ts

@@ -21,6 +21,7 @@ import {
   preProcessImageAndWebReferenceContent,
   preProcessImageContent,
   stream,
+  streamWithThink,
 } from "@/app/utils/chat";
 import { cloudflareAIGatewayUrl } from "@/app/utils/cloudflare";
 import { DalleSize, DalleQuality, DalleStyle } from "@/app/typing";
@@ -49,6 +50,7 @@ import {
   isVisionModel,
   isDalle3 as _isDalle3,
   getWebReferenceMessageTextContent,
+  getTimeoutMSByModel,
 } from "@/app/utils";
 
 export interface OpenAIListModelResponse {
@@ -73,6 +75,7 @@ export interface RequestPayload {
   frequency_penalty: number;
   top_p: number;
   max_tokens?: number;
+  max_completion_tokens?: number;
 }
 
 export interface DalleRequestPayload {
@@ -223,7 +226,9 @@ export class ChatGPTApi implements LLMApi {
     let requestPayload: RequestPayload | DalleRequestPayload;
 
     const isDalle3 = _isDalle3(options.config.model);
-    const isO1 = options.config.model.startsWith("o1");
+    const isO1OrO3 =
+      options.config.model.startsWith("o1") ||
+      options.config.model.startsWith("o3");
     if (isDalle3) {
       const prompt = getMessageTextContent(
         options.messages.slice(-1)?.pop() as any,
@@ -245,23 +250,28 @@ export class ChatGPTApi implements LLMApi {
         const content = visionModel
           ? await preProcessImageAndWebReferenceContent(v)
           : getWebReferenceMessageTextContent(v);
-        if (!(isO1 && v.role === "system"))
+        if (!(isO1OrO3 && v.role === "system"))
           messages.push({ role: v.role, content });
       }
 
       // O1 not support image, tools (plugin in ChatGPTNextWeb) and system, stream, logprobs, temperature, top_p, n, presence_penalty, frequency_penalty yet.
       requestPayload = {
         messages,
-        stream: !isO1 ? options.config.stream : false,
+        stream: options.config.stream,
         model: modelConfig.model,
-        temperature: !isO1 ? modelConfig.temperature : 1,
-        presence_penalty: !isO1 ? modelConfig.presence_penalty : 0,
-        frequency_penalty: !isO1 ? modelConfig.frequency_penalty : 0,
-        top_p: !isO1 ? modelConfig.top_p : 1,
+        temperature: !isO1OrO3 ? modelConfig.temperature : 1,
+        presence_penalty: !isO1OrO3 ? modelConfig.presence_penalty : 0,
+        frequency_penalty: !isO1OrO3 ? modelConfig.frequency_penalty : 0,
+        top_p: !isO1OrO3 ? modelConfig.top_p : 1,
         // max_tokens: Math.max(modelConfig.max_tokens, 1024),
         // Please do not ask me why not send max_tokens, no reason, this param is just shit, I dont want to explain anymore.
       };
 
+      // O1 使用 max_completion_tokens 控制token数 (https://platform.openai.com/docs/guides/reasoning#controlling-costs)
+      if (isO1OrO3) {
+        requestPayload["max_completion_tokens"] = modelConfig.max_tokens;
+      }
+
       // add max_tokens to vision model
       if (visionModel) {
         requestPayload["max_tokens"] = Math.max(modelConfig.max_tokens, 4000);
@@ -270,7 +280,7 @@ export class ChatGPTApi implements LLMApi {
 
     console.log("[Request] openai payload: ", requestPayload);
 
-    const shouldStream = !isDalle3 && !!options.config.stream && !isO1;
+    const shouldStream = !isDalle3 && !!options.config.stream;
     const controller = new AbortController();
     options.onController?.(controller);
 
@@ -307,15 +317,16 @@ export class ChatGPTApi implements LLMApi {
         );
       }
       if (shouldStream) {
+        let index = -1;
         // const [tools, funcs] = usePluginStore
         //   .getState()
         //   .getAsTools(
         //     useChatStore.getState().currentSession().mask?.plugin || [],
         //   );
-        // console.log("getAsTools", tools, funcs);
         const tools = null;
         const funcs: Record<string, Function> = {};
-        stream(
+        // console.log("getAsTools", tools, funcs);
+        streamWithThink(
           chatPath,
           requestPayload,
           getHeaders(),
@@ -330,14 +341,18 @@ export class ChatGPTApi implements LLMApi {
               delta: {
                 content: string;
                 tool_calls: ChatMessageTool[];
+                reasoning_content: string | null;
               };
             }>;
+
+            if (!choices?.length) return { isThinking: false, content: "" };
+
             const tool_calls = choices[0]?.delta?.tool_calls;
             if (tool_calls?.length > 0) {
-              const index = tool_calls[0]?.index;
               const id = tool_calls[0]?.id;
               const args = tool_calls[0]?.function?.arguments;
               if (id) {
+                index += 1;
                 runTools.push({
                   id,
                   type: tool_calls[0]?.type,
@@ -351,7 +366,37 @@ export class ChatGPTApi implements LLMApi {
                 runTools[index]["function"]["arguments"] += args;
               }
             }
-            return choices[0]?.delta?.content;
+
+            const reasoning = choices[0]?.delta?.reasoning_content;
+            const content = choices[0]?.delta?.content;
+
+            // Skip if both content and reasoning_content are empty or null
+            if (
+              (!reasoning || reasoning.length === 0) &&
+              (!content || content.length === 0)
+            ) {
+              return {
+                isThinking: false,
+                content: "",
+              };
+            }
+
+            if (reasoning && reasoning.length > 0) {
+              return {
+                isThinking: true,
+                content: reasoning,
+              };
+            } else if (content && content.length > 0) {
+              return {
+                isThinking: false,
+                content: content,
+              };
+            }
+
+            return {
+              isThinking: false,
+              content: "",
+            };
           },
           // processToolMessage, include tool_calls message and tool call results
           (
@@ -359,6 +404,8 @@ export class ChatGPTApi implements LLMApi {
             toolCallMessage: any,
             toolCallResult: any[],
           ) => {
+            // reset index value
+            index = -1;
             // @ts-ignore
             requestPayload?.messages?.splice(
               // @ts-ignore
@@ -381,7 +428,7 @@ export class ChatGPTApi implements LLMApi {
         // make a fetch request
         const requestTimeoutId = setTimeout(
           () => controller.abort(),
-          isDalle3 || isO1 ? REQUEST_TIMEOUT_MS * 2 : REQUEST_TIMEOUT_MS, // dalle3 using b64_json is slow.
+          getTimeoutMSByModel(options.config.model),
         );
 
         const res = await fetch(chatPath, chatPayload);