fix: tool call history (#4576)

docSite/content/zh-cn/docs/development/upgrading/496.md

@@ -18,7 +18,7 @@ weight: 794
 2. 支持以 MCP SSE 协议创建工具。
 3. 批量执行节点支持交互节点，可实现每一轮循环都人工参与。
 4. 增加工作台二级菜单，合并工具箱。
-5. 增加 grok3、GPT4.1、Gemini2.5 模型系统配置。
+5. 增加 grok3、GPT4.1、o系列、Gemini2.5 模型系统配置。
 
 ## ⚙️ 优化
 
@@ -32,4 +32,5 @@ weight: 794
 ## 🐛 修复
 
 1. 修复子工作流包含交互节点时，未成功恢复子工作流所有数据。
-2. completion v1 接口，未接受 interactive 参数，导致 API 调用失败。
\ No newline at end of file
+2. completion v1 接口，未接受 interactive 参数，导致 API 调用失败。
+3. 连续工具调用，上下文截断异常
\ No newline at end of file

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

@@ -123,13 +123,13 @@
       "type": "llm"
     },
     {
-      "model": "o3-mini",
-      "name": "o3-mini",
+      "model": "o4-mini",
+      "name": "o4-mini",
       "maxContext": 200000,
       "maxResponse": 100000,
       "quoteMaxToken": 120000,
       "maxTemperature": null,
-      "vision": false,
+      "vision": true,
       "toolChoice": true,
       "functionCall": false,
       "defaultSystemChatPrompt": "",
@@ -140,25 +140,49 @@
       "usedInQueryExtension": true,
       "customExtractPrompt": "",
       "usedInToolCall": true,
-      "defaultConfig": {
-        "stream": false
+      "defaultConfig": {},
+      "fieldMap": {
+        "max_tokens": "max_completion_tokens"
       },
+      "type": "llm",
+      "showTopP": true,
+      "showStopSign": false
+    },
+    {
+      "model": "o3",
+      "name": "o3",
+      "maxContext": 200000,
+      "maxResponse": 100000,
+      "quoteMaxToken": 120000,
+      "maxTemperature": null,
+      "vision": true,
+      "toolChoice": true,
+      "functionCall": false,
+      "defaultSystemChatPrompt": "",
+      "datasetProcess": true,
+      "usedInClassify": true,
+      "customCQPrompt": "",
+      "usedInExtractFields": true,
+      "usedInQueryExtension": true,
+      "customExtractPrompt": "",
+      "usedInToolCall": true,
+      "defaultConfig": {},
       "fieldMap": {
         "max_tokens": "max_completion_tokens"
       },
       "type": "llm",
       "showTopP": true,
-      "showStopSign": true
+      "showStopSign": false
     },
     {
-      "model": "o1-mini",
-      "name": "o1-mini",
-      "maxContext": 128000,
-      "maxResponse": 4000,
+      "model": "o3-mini",
+      "name": "o3-mini",
+      "maxContext": 200000,
+      "maxResponse": 100000,
       "quoteMaxToken": 120000,
       "maxTemperature": null,
       "vision": false,
-      "toolChoice": false,
+      "toolChoice": true,
       "functionCall": false,
       "defaultSystemChatPrompt": "",
       "datasetProcess": true,
@@ -168,9 +192,7 @@
       "usedInQueryExtension": true,
       "customExtractPrompt": "",
       "usedInToolCall": true,
-      "defaultConfig": {
-        "stream": false
-      },
+      "defaultConfig": {},
       "fieldMap": {
         "max_tokens": "max_completion_tokens"
       },
@@ -196,9 +218,33 @@
       "usedInQueryExtension": true,
       "customExtractPrompt": "",
       "usedInToolCall": true,
-      "defaultConfig": {
-        "stream": false
+      "defaultConfig": {},
+      "fieldMap": {
+        "max_tokens": "max_completion_tokens"
       },
+      "type": "llm",
+      "showTopP": true,
+      "showStopSign": true
+    },
+    {
+      "model": "o1-mini",
+      "name": "o1-mini",
+      "maxContext": 128000,
+      "maxResponse": 4000,
+      "quoteMaxToken": 120000,
+      "maxTemperature": null,
+      "vision": false,
+      "toolChoice": false,
+      "functionCall": false,
+      "defaultSystemChatPrompt": "",
+      "datasetProcess": true,
+      "usedInClassify": true,
+      "customCQPrompt": "",
+      "usedInExtractFields": true,
+      "usedInQueryExtension": true,
+      "customExtractPrompt": "",
+      "usedInToolCall": true,
+      "defaultConfig": {},
       "fieldMap": {
         "max_tokens": "max_completion_tokens"
       },

packages/service/core/chat/utils.ts

@@ -36,36 +36,44 @@ export const filterGPTMessageByMaxContext = async ({
   const systemPrompts: ChatCompletionMessageParam[] = messages.slice(0, chatStartIndex);
   const chatPrompts: ChatCompletionMessageParam[] = messages.slice(chatStartIndex);
 
+  if (chatPrompts.length === 0) {
+    return systemPrompts;
+  }
+
   // reduce token of systemPrompt
   maxContext -= await countGptMessagesTokens(systemPrompts);
 
+  /* 截取时候保证一轮内容的完整性
+    1. user - assistant - user
+    2. user - assistant - tool
+    3. user - assistant - tool - tool - tool
+    3. user - assistant - tool - assistant - tool
+    4. user - assistant - assistant - tool - tool
+  */
   // Save the last chat prompt(question)
-  const question = chatPrompts.pop();
-  if (!question) {
-    return systemPrompts;
-  }
-  const chats: ChatCompletionMessageParam[] = [question];
+  let chats: ChatCompletionMessageParam[] = [];
+  let tmpChats: ChatCompletionMessageParam[] = [];
 
-  // 从后往前截取对话内容, 每次需要截取2个
-  while (1) {
-    const assistant = chatPrompts.pop();
-    const user = chatPrompts.pop();
-    if (!assistant || !user) {
+  // 从后往前截取对话内容, 每次到 user 则认为是一组完整信息
+  while (chatPrompts.length > 0) {
+    const lastMessage = chatPrompts.pop();
+    if (!lastMessage) {
       break;
     }
 
-    const tokens = await countGptMessagesTokens([assistant, user]);
-    maxContext -= tokens;
-    /* 整体 tokens 超出范围，截断  */
-    if (maxContext < 0) {
-      break;
-    }
-
-    chats.unshift(assistant);
-    chats.unshift(user);
+    // 遇到 user，说明到了一轮完整信息，可以开始判断是否需要保留
+    if (lastMessage.role === ChatCompletionRequestMessageRoleEnum.User) {
+      const tokens = await countGptMessagesTokens([lastMessage, ...tmpChats]);
+      maxContext -= tokens;
+      // 该轮信息整体 tokens 超出范围，这段数据不要了
+      if (maxContext < 0) {
+        break;
+      }
 
-    if (chatPrompts.length === 0) {
-      break;
+      chats = [lastMessage, ...tmpChats].concat(chats);
+      tmpChats = [];
+    } else {
+      tmpChats.unshift(lastMessage);
     }
   }
 

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

@@ -285,7 +285,7 @@ export const runToolWithToolChoice = async (
     },
     toolModel
   );
-  // console.log(JSON.stringify(requestMessages, null, 2), '==requestBody');
+  // console.log(JSON.stringify(filterMessages, null, 2), '==requestMessages');
   /* Run llm */
   const {
     response: aiResponse,