Merge pull request #6204 from bestsanmao/ali_bytedance_reasoning_content

add 3 type of reasoning_content support (+deepseek-r1@OpenAI @Alibaba @ByteDance), parse <think></think> from SSE

app/client/platforms/alibaba.ts

@@ -1,12 +1,13 @@
 "use client";
+import { ApiPath, Alibaba, ALIBABA_BASE_URL } from "@/app/constant";
 import {
-  ApiPath,
-  Alibaba,
-  ALIBABA_BASE_URL,
-  REQUEST_TIMEOUT_MS,
-} from "@/app/constant";
-import { useAccessStore, useAppConfig, useChatStore } from "@/app/store";
-
+  useAccessStore,
+  useAppConfig,
+  useChatStore,
+  ChatMessageTool,
+  usePluginStore,
+} from "@/app/store";
+import { streamWithThink } from "@/app/utils/chat";
 import {
   ChatOptions,
   getHeaders,
@@ -15,14 +16,12 @@ import {
   SpeechOptions,
   MultimodalContent,
 } from "../api";
-import Locale from "../../locales";
-import {
-  EventStreamContentType,
-  fetchEventSource,
-} from "@fortaine/fetch-event-source";
-import { prettyObject } from "@/app/utils/format";
 import { getClientConfig } from "@/app/config/client";
-import { getMessageTextContent } from "@/app/utils";
+import {
+  getMessageTextContent,
+  getMessageTextContentWithoutThinking,
+  getTimeoutMSByModel,
+} from "@/app/utils";
 import { fetch } from "@/app/utils/stream";
 
 export interface OpenAIListModelResponse {
@@ -92,7 +91,10 @@ export class QwenApi implements LLMApi {
   async chat(options: ChatOptions) {
     const messages = options.messages.map((v) => ({
       role: v.role,
-      content: getMessageTextContent(v),
+      content:
+        v.role === "assistant"
+          ? getMessageTextContentWithoutThinking(v)
+          : getMessageTextContent(v),
     }));
 
     const modelConfig = {
@@ -122,134 +124,118 @@ export class QwenApi implements LLMApi {
     options.onController?.(controller);
 
     try {
+      const headers = {
+        ...getHeaders(),
+        "X-DashScope-SSE": shouldStream ? "enable" : "disable",
+      };
+
       const chatPath = this.path(Alibaba.ChatPath);
       const chatPayload = {
         method: "POST",
         body: JSON.stringify(requestPayload),
         signal: controller.signal,
-        headers: {
-          ...getHeaders(),
-          "X-DashScope-SSE": shouldStream ? "enable" : "disable",
-        },
+        headers: headers,
       };
 
       // make a fetch request
       const requestTimeoutId = setTimeout(
         () => controller.abort(),
-        REQUEST_TIMEOUT_MS,
+        getTimeoutMSByModel(options.config.model),
       );
 
       if (shouldStream) {
-        let responseText = "";
-        let remainText = "";
-        let finished = false;
-        let responseRes: Response;
-
-        // animate response to make it looks smooth
-        function animateResponseText() {
-          if (finished || controller.signal.aborted) {
-            responseText += remainText;
-            console.log("[Response Animation] finished");
-            if (responseText?.length === 0) {
-              options.onError?.(new Error("empty response from server"));
+        const [tools, funcs] = usePluginStore
+          .getState()
+          .getAsTools(
+            useChatStore.getState().currentSession().mask?.plugin || [],
+          );
+        return streamWithThink(
+          chatPath,
+          requestPayload,
+          headers,
+          tools as any,
+          funcs,
+          controller,
+          // parseSSE
+          (text: string, runTools: ChatMessageTool[]) => {
+            // console.log("parseSSE", text, runTools);
+            const json = JSON.parse(text);
+            const choices = json.output.choices as Array<{
+              message: {
+                content: string | null;
+                tool_calls: ChatMessageTool[];
+                reasoning_content: string | null;
+              };
+            }>;
+
+            if (!choices?.length) return { isThinking: false, content: "" };
+
+            const tool_calls = choices[0]?.message?.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) {
+                runTools.push({
+                  id,
+                  type: tool_calls[0]?.type,
+                  function: {
+                    name: tool_calls[0]?.function?.name as string,
+                    arguments: args,
+                  },
+                });
+              } else {
+                // @ts-ignore
+                runTools[index]["function"]["arguments"] += args;
+              }
             }
-            return;
-          }
-
-          if (remainText.length > 0) {
-            const fetchCount = Math.max(1, Math.round(remainText.length / 60));
-            const fetchText = remainText.slice(0, fetchCount);
-            responseText += fetchText;
-            remainText = remainText.slice(fetchCount);
-            options.onUpdate?.(responseText, fetchText);
-          }
-
-          requestAnimationFrame(animateResponseText);
-        }
-
-        // start animaion
-        animateResponseText();
-
-        const finish = () => {
-          if (!finished) {
-            finished = true;
-            options.onFinish(responseText + remainText, responseRes);
-          }
-        };
-
-        controller.signal.onabort = finish;
-
-        fetchEventSource(chatPath, {
-          fetch: fetch as any,
-          ...chatPayload,
-          async onopen(res) {
-            clearTimeout(requestTimeoutId);
-            const contentType = res.headers.get("content-type");
-            console.log(
-              "[Alibaba] request response content type: ",
-              contentType,
-            );
-            responseRes = res;
 
-            if (contentType?.startsWith("text/plain")) {
-              responseText = await res.clone().text();
-              return finish();
-            }
+            const reasoning = choices[0]?.message?.reasoning_content;
+            const content = choices[0]?.message?.content;
 
+            // Skip if both content and reasoning_content are empty or null
             if (
-              !res.ok ||
-              !res.headers
-                .get("content-type")
-                ?.startsWith(EventStreamContentType) ||
-              res.status !== 200
+              (!reasoning || reasoning.length === 0) &&
+              (!content || content.length === 0)
             ) {
-              const responseTexts = [responseText];
-              let extraInfo = await res.clone().text();
-              try {
-                const resJson = await res.clone().json();
-                extraInfo = prettyObject(resJson);
-              } catch {}
-
-              if (res.status === 401) {
-                responseTexts.push(Locale.Error.Unauthorized);
-              }
-
-              if (extraInfo) {
-                responseTexts.push(extraInfo);
-              }
-
-              responseText = responseTexts.join("\n\n");
-
-              return finish();
+              return {
+                isThinking: false,
+                content: "",
+              };
             }
-          },
-          onmessage(msg) {
-            if (msg.data === "[DONE]" || finished) {
-              return finish();
-            }
-            const text = msg.data;
-            try {
-              const json = JSON.parse(text);
-              const choices = json.output.choices as Array<{
-                message: { content: string };
-              }>;
-              const delta = choices[0]?.message?.content;
-              if (delta) {
-                remainText += delta;
-              }
-            } catch (e) {
-              console.error("[Request] parse error", text, msg);
+
+            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: "",
+            };
           },
-          onclose() {
-            finish();
-          },
-          onerror(e) {
-            options.onError?.(e);
-            throw e;
+          // processToolMessage, include tool_calls message and tool call results
+          (
+            requestPayload: RequestPayload,
+            toolCallMessage: any,
+            toolCallResult: any[],
+          ) => {
+            requestPayload?.input?.messages?.splice(
+              requestPayload?.input?.messages?.length,
+              0,
+              toolCallMessage,
+              ...toolCallResult,
+            );
           },
-          openWhenHidden: true,
-        });
+          options,
+        );
       } else {
         const res = await fetch(chatPath, chatPayload);
         clearTimeout(requestTimeoutId);

app/client/platforms/baidu.ts

@@ -1,10 +1,5 @@
 "use client";
-import {
-  ApiPath,
-  Baidu,
-  BAIDU_BASE_URL,
-  REQUEST_TIMEOUT_MS,
-} from "@/app/constant";
+import { ApiPath, Baidu, BAIDU_BASE_URL } from "@/app/constant";
 import { useAccessStore, useAppConfig, useChatStore } from "@/app/store";
 import { getAccessToken } from "@/app/utils/baidu";
 
@@ -23,7 +18,7 @@ import {
 } from "@fortaine/fetch-event-source";
 import { prettyObject } from "@/app/utils/format";
 import { getClientConfig } from "@/app/config/client";
-import { getMessageTextContent } from "@/app/utils";
+import { getMessageTextContent, getTimeoutMSByModel } from "@/app/utils";
 import { fetch } from "@/app/utils/stream";
 
 export interface OpenAIListModelResponse {
@@ -155,7 +150,7 @@ export class ErnieApi implements LLMApi {
       // make a fetch request
       const requestTimeoutId = setTimeout(
         () => controller.abort(),
-        REQUEST_TIMEOUT_MS,
+        getTimeoutMSByModel(options.config.model),
       );
 
       if (shouldStream) {

app/client/platforms/bytedance.ts

@@ -1,11 +1,12 @@
 "use client";
+import { ApiPath, ByteDance, BYTEDANCE_BASE_URL } from "@/app/constant";
 import {
-  ApiPath,
-  ByteDance,
-  BYTEDANCE_BASE_URL,
-  REQUEST_TIMEOUT_MS,
-} from "@/app/constant";
-import { useAccessStore, useAppConfig, useChatStore } from "@/app/store";
+  useAccessStore,
+  useAppConfig,
+  useChatStore,
+  ChatMessageTool,
+  usePluginStore,
+} from "@/app/store";
 
 import {
   ChatOptions,
@@ -15,14 +16,14 @@ import {
   MultimodalContent,
   SpeechOptions,
 } from "../api";
-import Locale from "../../locales";
-import {
-  EventStreamContentType,
-  fetchEventSource,
-} from "@fortaine/fetch-event-source";
-import { prettyObject } from "@/app/utils/format";
+
+import { streamWithThink } from "@/app/utils/chat";
 import { getClientConfig } from "@/app/config/client";
 import { preProcessImageContent } from "@/app/utils/chat";
+import {
+  getMessageTextContentWithoutThinking,
+  getTimeoutMSByModel,
+} from "@/app/utils";
 import { fetch } from "@/app/utils/stream";
 
 export interface OpenAIListModelResponse {
@@ -34,7 +35,7 @@ export interface OpenAIListModelResponse {
   }>;
 }
 
-interface RequestPayload {
+interface RequestPayloadForByteDance {
   messages: {
     role: "system" | "user" | "assistant";
     content: string | MultimodalContent[];
@@ -86,7 +87,10 @@ export class DoubaoApi implements LLMApi {
   async chat(options: ChatOptions) {
     const messages: ChatOptions["messages"] = [];
     for (const v of options.messages) {
-      const content = await preProcessImageContent(v.content);
+      const content =
+        v.role === "assistant"
+          ? getMessageTextContentWithoutThinking(v)
+          : await preProcessImageContent(v.content);
       messages.push({ role: v.role, content });
     }
 
@@ -99,7 +103,7 @@ export class DoubaoApi implements LLMApi {
     };
 
     const shouldStream = !!options.config.stream;
-    const requestPayload: RequestPayload = {
+    const requestPayload: RequestPayloadForByteDance = {
       messages,
       stream: shouldStream,
       model: modelConfig.model,
@@ -124,119 +128,101 @@ export class DoubaoApi implements LLMApi {
       // make a fetch request
       const requestTimeoutId = setTimeout(
         () => controller.abort(),
-        REQUEST_TIMEOUT_MS,
+        getTimeoutMSByModel(options.config.model),
       );
 
       if (shouldStream) {
-        let responseText = "";
-        let remainText = "";
-        let finished = false;
-        let responseRes: Response;
-
-        // animate response to make it looks smooth
-        function animateResponseText() {
-          if (finished || controller.signal.aborted) {
-            responseText += remainText;
-            console.log("[Response Animation] finished");
-            if (responseText?.length === 0) {
-              options.onError?.(new Error("empty response from server"));
-            }
-            return;
-          }
-
-          if (remainText.length > 0) {
-            const fetchCount = Math.max(1, Math.round(remainText.length / 60));
-            const fetchText = remainText.slice(0, fetchCount);
-            responseText += fetchText;
-            remainText = remainText.slice(fetchCount);
-            options.onUpdate?.(responseText, fetchText);
-          }
-
-          requestAnimationFrame(animateResponseText);
-        }
-
-        // start animaion
-        animateResponseText();
-
-        const finish = () => {
-          if (!finished) {
-            finished = true;
-            options.onFinish(responseText + remainText, responseRes);
-          }
-        };
-
-        controller.signal.onabort = finish;
-
-        fetchEventSource(chatPath, {
-          fetch: fetch as any,
-          ...chatPayload,
-          async onopen(res) {
-            clearTimeout(requestTimeoutId);
-            const contentType = res.headers.get("content-type");
-            console.log(
-              "[ByteDance] request response content type: ",
-              contentType,
-            );
-            responseRes = res;
-            if (contentType?.startsWith("text/plain")) {
-              responseText = await res.clone().text();
-              return finish();
+        const [tools, funcs] = usePluginStore
+          .getState()
+          .getAsTools(
+            useChatStore.getState().currentSession().mask?.plugin || [],
+          );
+        return streamWithThink(
+          chatPath,
+          requestPayload,
+          getHeaders(),
+          tools as any,
+          funcs,
+          controller,
+          // parseSSE
+          (text: string, runTools: ChatMessageTool[]) => {
+            // console.log("parseSSE", text, runTools);
+            const json = JSON.parse(text);
+            const choices = json.choices as Array<{
+              delta: {
+                content: string | null;
+                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) {
+                runTools.push({
+                  id,
+                  type: tool_calls[0]?.type,
+                  function: {
+                    name: tool_calls[0]?.function?.name as string,
+                    arguments: args,
+                  },
+                });
+              } else {
+                // @ts-ignore
+                runTools[index]["function"]["arguments"] += args;
+              }
             }
+            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 (
-              !res.ok ||
-              !res.headers
-                .get("content-type")
-                ?.startsWith(EventStreamContentType) ||
-              res.status !== 200
+              (!reasoning || reasoning.length === 0) &&
+              (!content || content.length === 0)
             ) {
-              const responseTexts = [responseText];
-              let extraInfo = await res.clone().text();
-              try {
-                const resJson = await res.clone().json();
-                extraInfo = prettyObject(resJson);
-              } catch {}
-
-              if (res.status === 401) {
-                responseTexts.push(Locale.Error.Unauthorized);
-              }
-
-              if (extraInfo) {
-                responseTexts.push(extraInfo);
-              }
-
-              responseText = responseTexts.join("\n\n");
-
-              return finish();
+              return {
+                isThinking: false,
+                content: "",
+              };
             }
-          },
-          onmessage(msg) {
-            if (msg.data === "[DONE]" || finished) {
-              return finish();
-            }
-            const text = msg.data;
-            try {
-              const json = JSON.parse(text);
-              const choices = json.choices as Array<{
-                delta: { content: string };
-              }>;
-              const delta = choices[0]?.delta?.content;
-              if (delta) {
-                remainText += delta;
-              }
-            } catch (e) {
-              console.error("[Request] parse error", text, msg);
+
+            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: "",
+            };
           },
-          onclose() {
-            finish();
-          },
-          onerror(e) {
-            options.onError?.(e);
-            throw e;
+          // processToolMessage, include tool_calls message and tool call results
+          (
+            requestPayload: RequestPayloadForByteDance,
+            toolCallMessage: any,
+            toolCallResult: any[],
+          ) => {
+            requestPayload?.messages?.splice(
+              requestPayload?.messages?.length,
+              0,
+              toolCallMessage,
+              ...toolCallResult,
+            );
           },
-          openWhenHidden: true,
-        });
+          options,
+        );
       } else {
         const res = await fetch(chatPath, chatPayload);
         clearTimeout(requestTimeoutId);

app/client/platforms/deepseek.ts

@@ -1,12 +1,6 @@
 "use client";
 // azure and openai, using same models. so using same LLMApi.
-import {
-  ApiPath,
-  DEEPSEEK_BASE_URL,
-  DeepSeek,
-  REQUEST_TIMEOUT_MS,
-  REQUEST_TIMEOUT_MS_FOR_THINKING,
-} from "@/app/constant";
+import { ApiPath, DEEPSEEK_BASE_URL, DeepSeek } from "@/app/constant";
 import {
   useAccessStore,
   useAppConfig,
@@ -26,6 +20,7 @@ import { getClientConfig } from "@/app/config/client";
 import {
   getMessageTextContent,
   getMessageTextContentWithoutThinking,
+  getTimeoutMSByModel,
 } from "@/app/utils";
 import { RequestPayload } from "./openai";
 import { fetch } from "@/app/utils/stream";
@@ -116,16 +111,10 @@ export class DeepSeekApi implements LLMApi {
         headers: getHeaders(),
       };
 
-      // console.log(chatPayload);
-
-      const isR1 =
-        options.config.model.endsWith("-reasoner") ||
-        options.config.model.endsWith("-r1");
-
       // make a fetch request
       const requestTimeoutId = setTimeout(
         () => controller.abort(),
-        isR1 ? REQUEST_TIMEOUT_MS_FOR_THINKING : REQUEST_TIMEOUT_MS,
+        getTimeoutMSByModel(options.config.model),
       );
 
       if (shouldStream) {
@@ -176,8 +165,8 @@ export class DeepSeekApi implements LLMApi {
 
             // Skip if both content and reasoning_content are empty or null
             if (
-              (!reasoning || reasoning.trim().length === 0) &&
-              (!content || content.trim().length === 0)
+              (!reasoning || reasoning.length === 0) &&
+              (!content || content.length === 0)
             ) {
               return {
                 isThinking: false,
@@ -185,12 +174,12 @@ export class DeepSeekApi implements LLMApi {
               };
             }
 
-            if (reasoning && reasoning.trim().length > 0) {
+            if (reasoning && reasoning.length > 0) {
               return {
                 isThinking: true,
                 content: reasoning,
               };
-            } else if (content && content.trim().length > 0) {
+            } else if (content && content.length > 0) {
               return {
                 isThinking: false,
                 content: content,

app/client/platforms/glm.ts

@@ -1,10 +1,5 @@
 "use client";
-import {
-  ApiPath,
-  CHATGLM_BASE_URL,
-  ChatGLM,
-  REQUEST_TIMEOUT_MS,
-} from "@/app/constant";
+import { ApiPath, CHATGLM_BASE_URL, ChatGLM } from "@/app/constant";
 import {
   useAccessStore,
   useAppConfig,
@@ -21,7 +16,11 @@ import {
   SpeechOptions,
 } from "../api";
 import { getClientConfig } from "@/app/config/client";
-import { getMessageTextContent, isVisionModel } from "@/app/utils";
+import {
+  getMessageTextContent,
+  isVisionModel,
+  getTimeoutMSByModel,
+} from "@/app/utils";
 import { RequestPayload } from "./openai";
 import { fetch } from "@/app/utils/stream";
 import { preProcessImageContent } from "@/app/utils/chat";
@@ -191,7 +190,7 @@ export class ChatGLMApi implements LLMApi {
 
       const requestTimeoutId = setTimeout(
         () => controller.abort(),
-        REQUEST_TIMEOUT_MS,
+        getTimeoutMSByModel(options.config.model),
       );
 
       if (modelType === "image" || modelType === "video") {

app/client/platforms/google.ts

@@ -1,9 +1,4 @@
-import {
-  ApiPath,
-  Google,
-  REQUEST_TIMEOUT_MS,
-  REQUEST_TIMEOUT_MS_FOR_THINKING,
-} from "@/app/constant";
+import { ApiPath, Google } from "@/app/constant";
 import {
   ChatOptions,
   getHeaders,
@@ -27,6 +22,7 @@ import {
   getMessageTextContent,
   getMessageImages,
   isVisionModel,
+  getTimeoutMSByModel,
 } from "@/app/utils";
 import { preProcessImageContent } from "@/app/utils/chat";
 import { nanoid } from "nanoid";
@@ -206,7 +202,7 @@ export class GeminiProApi implements LLMApi {
       // make a fetch request
       const requestTimeoutId = setTimeout(
         () => controller.abort(),
-        isThinking ? REQUEST_TIMEOUT_MS_FOR_THINKING : REQUEST_TIMEOUT_MS,
+        getTimeoutMSByModel(options.config.model),
       );
 
       if (shouldStream) {

app/client/platforms/openai.ts

@@ -8,7 +8,6 @@ import {
   Azure,
   REQUEST_TIMEOUT_MS,
   ServiceProvider,
-  REQUEST_TIMEOUT_MS_FOR_THINKING,
 } from "@/app/constant";
 import {
   ChatMessageTool,
@@ -22,7 +21,7 @@ import {
   preProcessImageContent,
   uploadImage,
   base64Image2Blob,
-  stream,
+  streamWithThink,
 } from "@/app/utils/chat";
 import { cloudflareAIGatewayUrl } from "@/app/utils/cloudflare";
 import { ModelSize, DalleQuality, DalleStyle } from "@/app/typing";
@@ -42,6 +41,7 @@ import {
   getMessageTextContent,
   isVisionModel,
   isDalle3 as _isDalle3,
+  getTimeoutMSByModel,
 } from "@/app/utils";
 import { fetch } from "@/app/utils/stream";
 
@@ -294,7 +294,7 @@ export class ChatGPTApi implements LLMApi {
             useChatStore.getState().currentSession().mask?.plugin || [],
           );
         // console.log("getAsTools", tools, funcs);
-        stream(
+        streamWithThink(
           chatPath,
           requestPayload,
           getHeaders(),
@@ -309,8 +309,12 @@ 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 id = tool_calls[0]?.id;
@@ -330,7 +334,37 @@ 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
           (
@@ -362,9 +396,7 @@ export class ChatGPTApi implements LLMApi {
         // make a fetch request
         const requestTimeoutId = setTimeout(
           () => controller.abort(),
-          isDalle3 || isO1OrO3
-            ? REQUEST_TIMEOUT_MS_FOR_THINKING
-            : REQUEST_TIMEOUT_MS, // dalle3 using b64_json is slow.
+          getTimeoutMSByModel(options.config.model),
         );
 
         const res = await fetch(chatPath, chatPayload);

app/client/platforms/siliconflow.ts

@@ -4,7 +4,6 @@ import {
   ApiPath,
   SILICONFLOW_BASE_URL,
   SiliconFlow,
-  REQUEST_TIMEOUT_MS_FOR_THINKING,
   DEFAULT_MODELS,
 } from "@/app/constant";
 import {
@@ -27,6 +26,7 @@ import {
   getMessageTextContent,
   getMessageTextContentWithoutThinking,
   isVisionModel,
+  getTimeoutMSByModel,
 } from "@/app/utils";
 import { RequestPayload } from "./openai";
 
@@ -137,7 +137,7 @@ export class SiliconflowApi implements LLMApi {
       // Use extended timeout for thinking models as they typically require more processing time
       const requestTimeoutId = setTimeout(
         () => controller.abort(),
-        REQUEST_TIMEOUT_MS_FOR_THINKING,
+        getTimeoutMSByModel(options.config.model),
       );
 
       if (shouldStream) {

app/client/platforms/tencent.ts

@@ -1,5 +1,5 @@
 "use client";
-import { ApiPath, TENCENT_BASE_URL, REQUEST_TIMEOUT_MS } from "@/app/constant";
+import { ApiPath, TENCENT_BASE_URL } from "@/app/constant";
 import { useAccessStore, useAppConfig, useChatStore } from "@/app/store";
 
 import {
@@ -17,7 +17,11 @@ import {
 } from "@fortaine/fetch-event-source";
 import { prettyObject } from "@/app/utils/format";
 import { getClientConfig } from "@/app/config/client";
-import { getMessageTextContent, isVisionModel } from "@/app/utils";
+import {
+  getMessageTextContent,
+  isVisionModel,
+  getTimeoutMSByModel,
+} from "@/app/utils";
 import mapKeys from "lodash-es/mapKeys";
 import mapValues from "lodash-es/mapValues";
 import isArray from "lodash-es/isArray";
@@ -135,7 +139,7 @@ export class HunyuanApi implements LLMApi {
       // make a fetch request
       const requestTimeoutId = setTimeout(
         () => controller.abort(),
-        REQUEST_TIMEOUT_MS,
+        getTimeoutMSByModel(options.config.model),
       );
 
       if (shouldStream) {

app/client/platforms/xai.ts

@@ -1,6 +1,6 @@
 "use client";
 // azure and openai, using same models. so using same LLMApi.
-import { ApiPath, XAI_BASE_URL, XAI, REQUEST_TIMEOUT_MS } from "@/app/constant";
+import { ApiPath, XAI_BASE_URL, XAI } from "@/app/constant";
 import {
   useAccessStore,
   useAppConfig,
@@ -17,6 +17,7 @@ import {
   SpeechOptions,
 } from "../api";
 import { getClientConfig } from "@/app/config/client";
+import { getTimeoutMSByModel } from "@/app/utils";
 import { preProcessImageContent } from "@/app/utils/chat";
 import { RequestPayload } from "./openai";
 import { fetch } from "@/app/utils/stream";
@@ -103,7 +104,7 @@ export class XAIApi implements LLMApi {
       // make a fetch request
       const requestTimeoutId = setTimeout(
         () => controller.abort(),
-        REQUEST_TIMEOUT_MS,
+        getTimeoutMSByModel(options.config.model),
       );
 
       if (shouldStream) {

app/utils.ts

@@ -2,7 +2,11 @@ import { useEffect, useState } from "react";
 import { showToast } from "./components/ui-lib";
 import Locale from "./locales";
 import { RequestMessage } from "./client/api";
-import { ServiceProvider } from "./constant";
+import {
+  REQUEST_TIMEOUT_MS,
+  REQUEST_TIMEOUT_MS_FOR_THINKING,
+  ServiceProvider,
+} from "./constant";
 // import { fetch as tauriFetch, ResponseType } from "@tauri-apps/api/http";
 import { fetch as tauriStreamFetch } from "./utils/stream";
 import { VISION_MODEL_REGEXES, EXCLUDE_VISION_MODEL_REGEXES } from "./constant";
@@ -292,6 +296,20 @@ export function isDalle3(model: string) {
   return "dall-e-3" === model;
 }
 
+export function getTimeoutMSByModel(model: string) {
+  model = model.toLowerCase();
+  if (
+    model.startsWith("dall-e") ||
+    model.startsWith("dalle") ||
+    model.startsWith("o1") ||
+    model.startsWith("o3") ||
+    model.includes("deepseek-r") ||
+    model.includes("-thinking")
+  )
+    return REQUEST_TIMEOUT_MS_FOR_THINKING;
+  return REQUEST_TIMEOUT_MS;
+}
+
 export function getModelSizes(model: string): ModelSize[] {
   if (isDalle3(model)) {
     return ["1024x1024", "1792x1024", "1024x1792"];

app/utils/chat.ts

@@ -400,6 +400,7 @@ export function streamWithThink(
   let responseRes: Response;
   let isInThinkingMode = false;
   let lastIsThinking = false;
+  let lastIsThinkingTagged = false; //between <think> and </think> tags
 
   // animate response to make it looks smooth
   function animateResponseText() {
@@ -579,6 +580,23 @@ export function streamWithThink(
           if (!chunk?.content || chunk.content.length === 0) {
             return;
           }
+
+          // deal with <think> and </think> tags start
+          if (!chunk.isThinking) {
+            if (chunk.content.startsWith("<think>")) {
+              chunk.isThinking = true;
+              chunk.content = chunk.content.slice(7).trim();
+              lastIsThinkingTagged = true;
+            } else if (chunk.content.endsWith("</think>")) {
+              chunk.isThinking = false;
+              chunk.content = chunk.content.slice(0, -8).trim();
+              lastIsThinkingTagged = false;
+            } else if (lastIsThinkingTagged) {
+              chunk.isThinking = true;
+            }
+          }
+          // deal with <think> and </think> tags start
+
           // Check if thinking mode changed
           const isThinkingChanged = lastIsThinking !== chunk.isThinking;
           lastIsThinking = chunk.isThinking;