perf: max_token count;feat: support resoner output;fix: member scroll (#3681)

* perf: supplement assistant empty response

* check array

* perf: max_token count

* feat: support resoner output

* member scroll

* update provider order

* i18n

packages/service/common/file/image/controller.ts

@@ -40,7 +40,7 @@ export async function uploadMongoImg({
     expiredTime: forever ? undefined : addHours(new Date(), 1)
   });
 
-  return `${process.env.FE_DOMAIN || ''}${process.env.NEXT_PUBLIC_BASE_URL || ''}${imageBaseUrl}${String(_id)}.${extension}`;
+  return `${process.env.NEXT_PUBLIC_BASE_URL || ''}${imageBaseUrl}${String(_id)}.${extension}`;
 }
 
 const getIdFromPath = (path?: string) => {

packages/service/core/ai/config/provider/DeepSeek.json

@@ -27,8 +27,9 @@
       "maxContext": 64000,
       "maxResponse": 4096,
       "quoteMaxToken": 60000,
-      "maxTemperature": 1.5,
+      "maxTemperature": null,
       "vision": false,
+      "reasoning": true,
       "toolChoice": false,
       "functionCall": false,
       "defaultSystemChatPrompt": "",
@@ -39,11 +40,9 @@
       "usedInQueryExtension": true,
       "customExtractPrompt": "",
       "usedInToolCall": true,
-      "defaultConfig": {
-        "temperature": null
-      },
+      "defaultConfig": {},
       "fieldMap": {},
       "type": "llm"
     }
   ]
-}
+}

packages/service/core/ai/config/provider/OpenAI.json

@@ -50,10 +50,10 @@
       "maxContext": 128000,
       "maxResponse": 4000,
       "quoteMaxToken": 120000,
-      "maxTemperature": 1.2,
+      "maxTemperature": null,
       "vision": false,
       "toolChoice": false,
-      "functionCall": true,
+      "functionCall": false,
       "defaultSystemChatPrompt": "",
       "datasetProcess": true,
       "usedInClassify": true,
@@ -63,8 +63,10 @@
       "customExtractPrompt": "",
       "usedInToolCall": true,
       "defaultConfig": {
-        "temperature": 1,
-        "max_tokens": null
+        "stream": false
+      },
+      "fieldMap": {
+        "max_tokens": "max_completion_tokens"
       },
       "type": "llm"
     },
@@ -74,10 +76,10 @@
       "maxContext": 128000,
       "maxResponse": 4000,
       "quoteMaxToken": 120000,
-      "maxTemperature": 1.2,
+      "maxTemperature": null,
       "vision": false,
       "toolChoice": false,
-      "functionCall": true,
+      "functionCall": false,
       "defaultSystemChatPrompt": "",
       "datasetProcess": true,
       "usedInClassify": true,
@@ -87,10 +89,11 @@
       "customExtractPrompt": "",
       "usedInToolCall": true,
       "defaultConfig": {
-        "temperature": 1,
-        "max_tokens": null,
         "stream": false
       },
+      "fieldMap": {
+        "max_tokens": "max_completion_tokens"
+      },
       "type": "llm"
     },
     {
@@ -99,10 +102,10 @@
       "maxContext": 195000,
       "maxResponse": 8000,
       "quoteMaxToken": 120000,
-      "maxTemperature": 1.2,
+      "maxTemperature": null,
       "vision": false,
       "toolChoice": false,
-      "functionCall": true,
+      "functionCall": false,
       "defaultSystemChatPrompt": "",
       "datasetProcess": true,
       "usedInClassify": true,
@@ -112,10 +115,11 @@
       "customExtractPrompt": "",
       "usedInToolCall": true,
       "defaultConfig": {
-        "temperature": 1,
-        "max_tokens": null,
         "stream": false
       },
+      "fieldMap": {
+        "max_tokens": "max_completion_tokens"
+      },
       "type": "llm"
     },
     {

packages/service/core/ai/functions/queryExtension.ts

@@ -2,10 +2,12 @@ import { replaceVariable } from '@fastgpt/global/common/string/tools';
 import { createChatCompletion } from '../config';
 import { ChatItemType } from '@fastgpt/global/core/chat/type';
 import { countGptMessagesTokens, countPromptTokens } from '../../../common/string/tiktoken/index';
-import { chatValue2RuntimePrompt } from '@fastgpt/global/core/chat/adapt';
+import { chats2GPTMessages } from '@fastgpt/global/core/chat/adapt';
 import { getLLMModel } from '../model';
 import { llmCompletionsBodyFormat } from '../utils';
 import { addLog } from '../../../common/system/log';
+import { filterGPTMessageByMaxContext } from '../../chat/utils';
+import json5 from 'json5';
 
 /* 
     query extension - 问题扩展
@@ -13,72 +15,73 @@ import { addLog } from '../../../common/system/log';
 */
 
 const title = global.feConfigs?.systemTitle || 'FastAI';
-const defaultPrompt = `作为一个向量检索助手，你的任务是结合历史记录，从不同角度，为“原问题”生成个不同版本的“检索词”，从而提高向量检索的语义丰富度，提高向量检索的精度。
+const defaultPrompt = `## 你的任务
+你作为一个向量检索助手，你的任务是结合历史记录，从不同角度，为“原问题”生成个不同版本的“检索词”，从而提高向量检索的语义丰富度，提高向量检索的精度。
 生成的问题要求指向对象清晰明确，并与“原问题语言相同”。
 
-参考 <Example></Example> 标中的示例来完成任务。
+## 参考示例
 
-<Example>
 历史记录: 
 """
+null
 """
 原问题: 介绍下剧情。
 检索词: ["介绍下故事的背景。","故事的主题是什么？","介绍下故事的主要人物。"]
 ----------------
 历史记录: 
 """
-Q: 对话背景。
-A: 当前对话是关于 Nginx 的介绍和使用等。
+user: 对话背景。
+assistant: 当前对话是关于 Nginx 的介绍和使用等。
 """
 原问题: 怎么下载
 检索词: ["Nginx 如何下载？","下载 Nginx 需要什么条件？","有哪些渠道可以下载 Nginx？"]
 ----------------
 历史记录: 
 """
-Q: 对话背景。
-A: 当前对话是关于 Nginx 的介绍和使用等。
-Q: 报错 "no connection"
-A: 报错"no connection"可能是因为……
+user: 对话背景。
+assistant: 当前对话是关于 Nginx 的介绍和使用等。
+user: 报错 "no connection"
+assistant: 报错"no connection"可能是因为……
 """
 原问题: 怎么解决
 检索词: ["Nginx报错"no connection"如何解决？","造成'no connection'报错的原因。","Nginx提示'no connection'，要怎么办？"]
 ----------------
 历史记录: 
 """
-Q: 护产假多少天?
-A: 护产假的天数根据员工所在的城市而定。请提供您所在的城市，以便我回答您的问题。
+user: How long is the maternity leave?
+assistant: The number of days of maternity leave depends on the city in which the employee is located. Please provide your city so that I can answer your questions.
 """
-原问题: 沈阳
-检索词: ["沈阳的护产假多少天？","沈阳的护产假政策。","沈阳的护产假标准。"]
+原问题: ShenYang
+检索词: ["How many days is maternity leave in Shenyang?","Shenyang's maternity leave policy.","The standard of maternity leave in Shenyang."]
 ----------------
 历史记录: 
 """
-Q: 作者是谁？
-A: ${title} 的作者是 labring。
+user: 作者是谁？
+assistant: ${title} 的作者是 labring。
 """
 原问题: Tell me about him
 检索词: ["Introduce labring, the author of ${title}." ," Background information on author labring." "," Why does labring do ${title}?"]
 ----------------
 历史记录:
 """
-Q: 对话背景。
-A: 关于 ${title} 的介绍和使用等问题。
+user: 对话背景。
+assistant: 关于 ${title} 的介绍和使用等问题。
 """
 原问题: 你好。
 检索词: ["你好"]
 ----------------
 历史记录:
 """
-Q: ${title} 如何收费？
-A: ${title} 收费可以参考……
+user: ${title} 如何收费？
+assistant: ${title} 收费可以参考……
 """
 原问题: 你知道 laf 么？
 检索词: ["laf 的官网地址是多少？","laf 的使用教程。","laf 有什么特点和优势。"]
 ----------------
 历史记录:
 """
-Q: ${title} 的优势
-A: 1. 开源
+user: ${title} 的优势
+assistant: 1. 开源
    2. 简便
    3. 扩展性强
 """
@@ -87,18 +90,20 @@ A: 1. 开源
 ----------------
 历史记录:
 """
-Q: 什么是 ${title}？
-A: ${title} 是一个 RAG 平台。
-Q: 什么是 Laf？
-A: Laf 是一个云函数开发平台。
+user: 什么是 ${title}？
+assistant: ${title} 是一个 RAG 平台。
+user: 什么是 Laf？
+assistant: Laf 是一个云函数开发平台。
 """
 原问题: 它们有什么关系？
 检索词: ["${title}和Laf有什么关系？","介绍下${title}","介绍下Laf"]
-</Example>
 
------
+## 输出要求
 
-下面是正式的任务：
+1. 输出格式为 JSON 数组，数组中每个元素为字符串。无需对输出进行任何解释。
+2. 输出语言与原问题相同。原问题为中文则输出中文；原问题为英文则输出英文。
+
+## 开始任务
 
 历史记录:
 """
@@ -125,26 +130,39 @@ export const queryExtension = async ({
   outputTokens: number;
 }> => {
   const systemFewShot = chatBg
-    ? `Q: 对话背景。
-A: ${chatBg}
+    ? `user: 对话背景。
+assistant: ${chatBg}
 `
     : '';
-  const historyFewShot = histories
-    .map((item) => {
-      const role = item.obj === 'Human' ? 'Q' : 'A';
-      return `${role}: ${chatValue2RuntimePrompt(item.value).text}`;
-    })
-    .join('\n');
-  const concatFewShot = `${systemFewShot}${historyFewShot}`.trim();
 
   const modelData = getLLMModel(model);
+  const filterHistories = await filterGPTMessageByMaxContext({
+    messages: chats2GPTMessages({ messages: histories, reserveId: false }),
+    maxContext: modelData.maxContext - 1000
+  });
+
+  const historyFewShot = filterHistories
+    .map((item) => {
+      const role = item.role;
+      const content = item.content;
+      if ((role === 'user' || role === 'assistant') && content) {
+        if (typeof content === 'string') {
+          return `${role}: ${content}`;
+        } else {
+          return `${role}: ${content.map((item) => (item.type === 'text' ? item.text : '')).join('\n')}`;
+        }
+      }
+    })
+    .filter(Boolean)
+    .join('\n');
+  const concatFewShot = `${systemFewShot}${historyFewShot}`.trim();
 
   const messages = [
     {
       role: 'user',
       content: replaceVariable(defaultPrompt, {
         query: `${query}`,
-        histories: concatFewShot
+        histories: concatFewShot || 'null'
       })
     }
   ] as any;
@@ -154,7 +172,7 @@ A: ${chatBg}
       {
         stream: false,
         model: modelData.model,
-        temperature: 0.01,
+        temperature: 0.1,
         messages
       },
       modelData
@@ -172,22 +190,41 @@ A: ${chatBg}
     };
   }
 
+  const start = answer.indexOf('[');
+  const end = answer.lastIndexOf(']');
+  if (start === -1 || end === -1) {
+    addLog.warn('Query extension failed, not a valid JSON', {
+      answer
+    });
+    return {
+      rawQuery: query,
+      extensionQueries: [],
+      model,
+      inputTokens: 0,
+      outputTokens: 0
+    };
+  }
+
   // Intercept the content of [] and retain []
-  answer = answer.match(/\[.*?\]/)?.[0] || '';
-  answer = answer.replace(/\\"/g, '"');
+  const jsonStr = answer
+    .substring(start, end + 1)
+    .replace(/(\\n|\\)/g, '')
+    .replace(/  /g, '');
 
   try {
-    const queries = JSON.parse(answer) as string[];
+    const queries = json5.parse(jsonStr) as string[];
 
     return {
       rawQuery: query,
-      extensionQueries: Array.isArray(queries) ? queries : [],
+      extensionQueries: (Array.isArray(queries) ? queries : []).slice(0, 5),
       model,
       inputTokens: await countGptMessagesTokens(messages),
       outputTokens: await countPromptTokens(answer)
     };
   } catch (error) {
-    addLog.error(`Query extension error`, error);
+    addLog.warn('Query extension failed, not a valid JSON', {
+      answer
+    });
     return {
       rawQuery: query,
       extensionQueries: [],

packages/service/core/ai/utils.ts

@@ -2,33 +2,23 @@ import { LLMModelItemType } from '@fastgpt/global/core/ai/model.d';
 import {
   ChatCompletionCreateParamsNonStreaming,
   ChatCompletionCreateParamsStreaming,
-  ChatCompletionMessageParam,
   StreamChatType
 } from '@fastgpt/global/core/ai/type';
-import { countGptMessagesTokens } from '../../common/string/tiktoken';
 import { getLLMModel } from './model';
 
-export const computedMaxToken = async ({
+/* 
+  Count response max token
+*/
+export const computedMaxToken = ({
   maxToken,
-  model,
-  filterMessages = []
+  model
 }: {
   maxToken?: number;
   model: LLMModelItemType;
-  filterMessages: ChatCompletionMessageParam[];
 }) => {
   if (maxToken === undefined) return;
 
   maxToken = Math.min(maxToken, model.maxResponse);
-  const tokensLimit = model.maxContext;
-
-  /* count response max token */
-  const promptsToken = await countGptMessagesTokens(filterMessages);
-  maxToken = promptsToken + maxToken > tokensLimit ? tokensLimit - promptsToken : maxToken;
-
-  if (maxToken <= 0) {
-    maxToken = 200;
-  }
   return maxToken;
 };
 
@@ -40,6 +30,7 @@ export const computedTemperature = ({
   model: LLMModelItemType;
   temperature: number;
 }) => {
+  if (typeof model.maxTemperature !== 'number') return undefined;
   temperature = +(model.maxTemperature * (temperature / 10)).toFixed(2);
   temperature = Math.max(temperature, 0.01);
 

packages/service/core/chat/utils.ts

@@ -14,36 +14,19 @@ import { serverRequestBaseUrl } from '../../common/api/serverRequest';
 import { i18nT } from '../../../web/i18n/utils';
 import { addLog } from '../../common/system/log';
 
-export const filterGPTMessageByMaxTokens = async ({
+export const filterGPTMessageByMaxContext = async ({
   messages = [],
-  maxTokens
+  maxContext
 }: {
   messages: ChatCompletionMessageParam[];
-  maxTokens: number;
+  maxContext: number;
 }) => {
   if (!Array.isArray(messages)) {
     return [];
   }
-  const rawTextLen = messages.reduce((sum, item) => {
-    if (typeof item.content === 'string') {
-      return sum + item.content.length;
-    }
-    if (Array.isArray(item.content)) {
-      return (
-        sum +
-        item.content.reduce((sum, item) => {
-          if (item.type === 'text') {
-            return sum + item.text.length;
-          }
-          return sum;
-        }, 0)
-      );
-    }
-    return sum;
-  }, 0);
 
   // If the text length is less than half of the maximum token, no calculation is required
-  if (rawTextLen < maxTokens * 0.5) {
+  if (messages.length < 4) {
     return messages;
   }
 
@@ -55,7 +38,7 @@ export const filterGPTMessageByMaxTokens = async ({
   const chatPrompts: ChatCompletionMessageParam[] = messages.slice(chatStartIndex);
 
   // reduce token of systemPrompt
-  maxTokens -= await countGptMessagesTokens(systemPrompts);
+  maxContext -= await countGptMessagesTokens(systemPrompts);
 
   // Save the last chat prompt(question)
   const question = chatPrompts.pop();
@@ -73,9 +56,9 @@ export const filterGPTMessageByMaxTokens = async ({
     }
 
     const tokens = await countGptMessagesTokens([assistant, user]);
-    maxTokens -= tokens;
+    maxContext -= tokens;
     /* 整体 tokens 超出范围，截断  */
-    if (maxTokens < 0) {
+    if (maxContext < 0) {
       break;
     }
 

packages/service/core/workflow/dispatch/agent/extract.ts

@@ -1,5 +1,5 @@
 import { chats2GPTMessages } from '@fastgpt/global/core/chat/adapt';
-import { filterGPTMessageByMaxTokens, loadRequestMessages } from '../../../chat/utils';
+import { filterGPTMessageByMaxContext, loadRequestMessages } from '../../../chat/utils';
 import type { ChatItemType } from '@fastgpt/global/core/chat/type.d';
 import {
   countMessagesTokens,
@@ -175,9 +175,9 @@ ${description ? `- ${description}` : ''}
     }
   ];
   const adaptMessages = chats2GPTMessages({ messages, reserveId: false });
-  const filterMessages = await filterGPTMessageByMaxTokens({
+  const filterMessages = await filterGPTMessageByMaxContext({
     messages: adaptMessages,
-    maxTokens: extractModel.maxContext
+    maxContext: extractModel.maxContext
   });
   const requestMessages = await loadRequestMessages({
     messages: filterMessages,

packages/service/core/workflow/dispatch/agent/runTool/functionCall.ts

@@ -1,5 +1,5 @@
 import { createChatCompletion } from '../../../../ai/config';
-import { filterGPTMessageByMaxTokens, loadRequestMessages } from '../../../../chat/utils';
+import { filterGPTMessageByMaxContext, loadRequestMessages } from '../../../../chat/utils';
 import {
   ChatCompletion,
   StreamChatType,
@@ -172,10 +172,14 @@ export const runToolWithFunctionCall = async (
     };
   });
 
+  const max_tokens = computedMaxToken({
+    model: toolModel,
+    maxToken
+  });
   const filterMessages = (
-    await filterGPTMessageByMaxTokens({
+    await filterGPTMessageByMaxContext({
       messages,
-      maxTokens: toolModel.maxContext - 300 // filter token. not response maxToken
+      maxContext: toolModel.maxContext - (max_tokens || 0) // filter token. not response maxToken
     })
   ).map((item) => {
     if (item.role === ChatCompletionRequestMessageRoleEnum.Assistant && item.function_call) {
@@ -190,16 +194,11 @@ export const runToolWithFunctionCall = async (
     }
     return item;
   });
-  const [requestMessages, max_tokens] = await Promise.all([
+  const [requestMessages] = await Promise.all([
     loadRequestMessages({
       messages: filterMessages,
       useVision: toolModel.vision && aiChatVision,
       origin: requestOrigin
-    }),
-    computedMaxToken({
-      model: toolModel,
-      maxToken,
-      filterMessages
     })
   ]);
   const requestBody = llmCompletionsBodyFormat(

packages/service/core/workflow/dispatch/agent/runTool/promptCall.ts

@@ -1,5 +1,5 @@
 import { createChatCompletion } from '../../../../ai/config';
-import { filterGPTMessageByMaxTokens, loadRequestMessages } from '../../../../chat/utils';
+import { filterGPTMessageByMaxContext, loadRequestMessages } from '../../../../chat/utils';
 import {
   ChatCompletion,
   StreamChatType,
@@ -196,21 +196,20 @@ export const runToolWithPromptCall = async (
     return Promise.reject('Prompt call invalid input');
   }
 
-  const filterMessages = await filterGPTMessageByMaxTokens({
+  const max_tokens = computedMaxToken({
+    model: toolModel,
+    maxToken
+  });
+  const filterMessages = await filterGPTMessageByMaxContext({
     messages,
-    maxTokens: toolModel.maxContext - 500 // filter token. not response maxToken
+    maxContext: toolModel.maxContext - (max_tokens || 0) // filter token. not response maxToken
   });
 
-  const [requestMessages, max_tokens] = await Promise.all([
+  const [requestMessages] = await Promise.all([
     loadRequestMessages({
       messages: filterMessages,
       useVision: toolModel.vision && aiChatVision,
       origin: requestOrigin
-    }),
-    computedMaxToken({
-      model: toolModel,
-      maxToken,
-      filterMessages
     })
   ]);
   const requestBody = llmCompletionsBodyFormat(

packages/service/core/workflow/dispatch/agent/runTool/toolChoice.ts

@@ -1,5 +1,5 @@
 import { createChatCompletion } from '../../../../ai/config';
-import { filterGPTMessageByMaxTokens, loadRequestMessages } from '../../../../chat/utils';
+import { filterGPTMessageByMaxContext, loadRequestMessages } from '../../../../chat/utils';
 import {
   ChatCompletion,
   ChatCompletionMessageToolCall,
@@ -228,11 +228,16 @@ export const runToolWithToolChoice = async (
     };
   });
 
+  const max_tokens = computedMaxToken({
+    model: toolModel,
+    maxToken
+  });
+
   // Filter histories by maxToken
   const filterMessages = (
-    await filterGPTMessageByMaxTokens({
+    await filterGPTMessageByMaxContext({
       messages,
-      maxTokens: toolModel.maxContext - 300 // filter token. not response maxToken
+      maxContext: toolModel.maxContext - (max_tokens || 0) // filter token. not response maxToken
     })
   ).map((item) => {
     if (item.role === 'assistant' && item.tool_calls) {
@@ -248,16 +253,11 @@ export const runToolWithToolChoice = async (
     return item;
   });
 
-  const [requestMessages, max_tokens] = await Promise.all([
+  const [requestMessages] = await Promise.all([
     loadRequestMessages({
       messages: filterMessages,
       useVision: toolModel.vision && aiChatVision,
       origin: requestOrigin
-    }),
-    computedMaxToken({
-      model: toolModel,
-      maxToken,
-      filterMessages
     })
   ]);
   const requestBody = llmCompletionsBodyFormat(

packages/service/core/workflow/dispatch/chat/oneapi.ts

@@ -1,5 +1,5 @@
 import type { NextApiResponse } from 'next';
-import { filterGPTMessageByMaxTokens, loadRequestMessages } from '../../../chat/utils';
+import { filterGPTMessageByMaxContext, loadRequestMessages } from '../../../chat/utils';
 import type { ChatItemType, UserChatItemValueItemType } from '@fastgpt/global/core/chat/type.d';
 import { ChatRoleEnum } from '@fastgpt/global/core/chat/constants';
 import { SseResponseEventEnum } from '@fastgpt/global/core/workflow/runtime/constants';
@@ -58,6 +58,7 @@ export type ChatProps = ModuleDispatchProps<
 >;
 export type ChatResponse = DispatchNodeResultType<{
   [NodeOutputKeyEnum.answerText]: string;
+  [NodeOutputKeyEnum.reasoningText]?: string;
   [NodeOutputKeyEnum.history]: ChatItemType[];
 }>;
 
@@ -87,22 +88,24 @@ export const dispatchChatCompletion = async (props: ChatProps): Promise<ChatResp
       quoteTemplate,
       quotePrompt,
       aiChatVision,
+      aiChatReasoning,
       fileUrlList: fileLinks, // node quote file links
       stringQuoteText //abandon
     }
   } = props;
   const { files: inputFiles } = chatValue2RuntimePrompt(query); // Chat box input files
 
-  stream = stream && isResponseAnswerText;
-
-  const chatHistories = getHistories(history, histories);
-  quoteQA = checkQuoteQAValue(quoteQA);
-
   const modelConstantsData = getLLMModel(model);
   if (!modelConstantsData) {
     return Promise.reject('The chat model is undefined, you need to select a chat model.');
   }
 
+  stream = stream && isResponseAnswerText;
+  aiChatReasoning = !!aiChatReasoning && !!modelConstantsData.reasoning;
+
+  const chatHistories = getHistories(history, histories);
+  quoteQA = checkQuoteQAValue(quoteQA);
+
   const [{ datasetQuoteText }, { documentQuoteText, userFiles }] = await Promise.all([
     filterDatasetQuote({
       quoteQA,
@@ -124,9 +127,15 @@ export const dispatchChatCompletion = async (props: ChatProps): Promise<ChatResp
     return Promise.reject(i18nT('chat:AI_input_is_empty'));
   }
 
+  const max_tokens = computedMaxToken({
+    model: modelConstantsData,
+    maxToken
+  });
+
   const [{ filterMessages }] = await Promise.all([
     getChatMessages({
       model: modelConstantsData,
+      maxTokens: max_tokens,
       histories: chatHistories,
       useDatasetQuote: quoteQA !== undefined,
       datasetQuoteText,
@@ -137,8 +146,8 @@ export const dispatchChatCompletion = async (props: ChatProps): Promise<ChatResp
       userFiles,
       documentQuoteText
     }),
+    // Censor = true and system key, will check content
     (() => {
-      // censor model and system key
       if (modelConstantsData.censor && !externalProvider.openaiAccount?.key) {
         return postTextCensor({
           text: `${systemPrompt}
@@ -149,18 +158,11 @@ export const dispatchChatCompletion = async (props: ChatProps): Promise<ChatResp
     })()
   ]);
 
-  const [requestMessages, max_tokens] = await Promise.all([
-    loadRequestMessages({
-      messages: filterMessages,
-      useVision: modelConstantsData.vision && aiChatVision,
-      origin: requestOrigin
-    }),
-    computedMaxToken({
-      model: modelConstantsData,
-      maxToken,
-      filterMessages
-    })
-  ]);
+  const requestMessages = await loadRequestMessages({
+    messages: filterMessages,
+    useVision: modelConstantsData.vision && aiChatVision,
+    origin: requestOrigin
+  });
 
   const requestBody = llmCompletionsBodyFormat(
     {
@@ -183,34 +185,42 @@ export const dispatchChatCompletion = async (props: ChatProps): Promise<ChatResp
     }
   });
 
-  const { answerText } = await (async () => {
+  const { answerText, reasoningText } = await (async () => {
     if (res && isStreamResponse) {
       // sse response
-      const { answer } = await streamResponse({
+      const { answer, reasoning } = await streamResponse({
         res,
         stream: response,
+        aiChatReasoning,
         workflowStreamResponse
       });
 
       return {
-        answerText: answer
+        answerText: answer,
+        reasoningText: reasoning
       };
     } else {
       const unStreamResponse = response as ChatCompletion;
       const answer = unStreamResponse.choices?.[0]?.message?.content || '';
-
+      const reasoning = aiChatReasoning
+        ? // @ts-ignore
+          unStreamResponse.choices?.[0]?.message?.reasoning_content || ''
+        : '';
       if (stream) {
         // Some models do not support streaming
-        workflowStreamResponse?.({
-          event: SseResponseEventEnum.fastAnswer,
-          data: textAdaptGptResponse({
-            text: answer
-          })
-        });
+        reasoning &&
+          workflowStreamResponse?.({
+            event: SseResponseEventEnum.fastAnswer,
+            data: textAdaptGptResponse({
+              text: answer,
+              reasoning_content: reasoning
+            })
+          });
       }
 
       return {
-        answerText: answer
+        answerText: answer,
+        reasoningText: reasoning
       };
     }
   })();
@@ -241,6 +251,7 @@ export const dispatchChatCompletion = async (props: ChatProps): Promise<ChatResp
 
   return {
     answerText,
+    reasoningText,
     [DispatchNodeResponseKeyEnum.nodeResponse]: {
       totalPoints: externalProvider.openaiAccount?.key ? 0 : totalPoints,
       model: modelName,
@@ -367,6 +378,7 @@ async function getMultiInput({
 
 async function getChatMessages({
   model,
+  maxTokens = 0,
   aiChatQuoteRole,
   datasetQuotePrompt = '',
   datasetQuoteText,
@@ -378,6 +390,7 @@ async function getChatMessages({
   documentQuoteText
 }: {
   model: LLMModelItemType;
+  maxTokens?: number;
   // dataset quote
   aiChatQuoteRole: AiChatQuoteRoleType; // user: replace user prompt; system: replace system prompt
   datasetQuotePrompt?: string;
@@ -444,9 +457,9 @@ async function getChatMessages({
 
   const adaptMessages = chats2GPTMessages({ messages, reserveId: false });
 
-  const filterMessages = await filterGPTMessageByMaxTokens({
+  const filterMessages = await filterGPTMessageByMaxContext({
     messages: adaptMessages,
-    maxTokens: model.maxContext - 300 // filter token. not response maxToken
+    maxContext: model.maxContext - maxTokens // filter token. not response maxToken
   });
 
   return {
@@ -457,33 +470,43 @@ async function getChatMessages({
 async function streamResponse({
   res,
   stream,
-  workflowStreamResponse
+  workflowStreamResponse,
+  aiChatReasoning
 }: {
   res: NextApiResponse;
   stream: StreamChatType;
   workflowStreamResponse?: WorkflowResponseType;
+  aiChatReasoning?: boolean;
 }) {
   const write = responseWriteController({
     res,
     readStream: stream
   });
   let answer = '';
+  let reasoning = '';
   for await (const part of stream) {
     if (res.closed) {
       stream.controller?.abort();
       break;
     }
+
     const content = part.choices?.[0]?.delta?.content || '';
     answer += content;
 
+    const reasoningContent = aiChatReasoning
+      ? part.choices?.[0]?.delta?.reasoning_content || ''
+      : '';
+    reasoning += reasoningContent;
+
     workflowStreamResponse?.({
       write,
       event: SseResponseEventEnum.answer,
       data: textAdaptGptResponse({
-        text: content
+        text: content,
+        reasoning_content: reasoningContent
       })
     });
   }
 
-  return { answer };
+  return { answer, reasoning };
 }

packages/service/core/workflow/dispatch/index.ts

@@ -204,6 +204,7 @@ export async function dispatchWorkFlow(data: Props): Promise<DispatchFlowRespons
     { inputs = [] }: RuntimeNodeItemType,
     {
       answerText = '',
+      reasoningText,
       responseData,
       nodeDispatchUsages,
       toolResponses,
@@ -213,6 +214,7 @@ export async function dispatchWorkFlow(data: Props): Promise<DispatchFlowRespons
     }: Omit<
       DispatchNodeResultType<{
         [NodeOutputKeyEnum.answerText]?: string;
+        [NodeOutputKeyEnum.reasoningText]?: string;
         [DispatchNodeResponseKeyEnum.nodeResponse]?: ChatHistoryItemResType;
       }>,
       'nodeResponse'
@@ -251,6 +253,13 @@ export async function dispatchWorkFlow(data: Props): Promise<DispatchFlowRespons
           }
         });
       }
+    } else if (reasoningText) {
+      chatAssistantResponse.push({
+        type: ChatItemValueTypeEnum.reasoning,
+        reasoning: {
+          content: reasoningText
+        }
+      });
     }
 
     if (rewriteHistories) {