Merge pull request #219 from Chancemrli/feat/cloudbase-agent-error-parse

feat: 添加LLM返回错误的解析逻辑

Author: Xxxxie

cloudrunfunctions/cloudbase-agent/src/chat_main.service.ts

@@ -101,7 +101,7 @@ export class MainChatService {
 
   async afterStream ({ error, needSave, callMsg, chunks, recordId = '' }) {
     if (error) {
-      console.log('请求大模型错误:', error)
+      console.log('afterStream_error: 请求大模型失败:', JSON.stringify(error))
     }
     if (needSave && recordId !== '') {
       const newChatEntity = new ChatHistoryEntity()

cloudrunfunctions/cloudbase-agent/src/chat_wx.service.ts

@@ -14,8 +14,13 @@ import {
   BOT_ROLE_ASSISTANT,
   BOT_ROLE_USER,
   BOT_TYPE_TEXT,
+  EXCEED_CONCURRENT_REQUEST_LIMIT,
+  EXCEED_CONCURRENT_REQUEST_LIMIT_MESSAGE,
+  EXCEED_TOKEN_QUOTA_LIMIT,
+  EXCEED_TOKEN_QUOTA_LIMIT_MESSAGE,
   MSG_TYPE_TEXT,
   MSG_TYPE_VOICE,
+  REQUEST_LLM_ERROR_MESSAGE,
   TRIGGER_SRC_WX_CUSTOM_SERVICE,
   TRIGGER_SRC_WX_MINI_APP,
   TRIGGER_SRC_WX_SERVICE,
@@ -471,7 +476,18 @@ export class WxChatService {
         }
       })
 
-      replyMsgData.content = result.content
+      if (result.error) {
+        // LLM 调用失败，映射错误码到用户友好消息
+        let errMsg = REQUEST_LLM_ERROR_MESSAGE
+        if (result.error.code === EXCEED_CONCURRENT_REQUEST_LIMIT) {
+          errMsg = EXCEED_CONCURRENT_REQUEST_LIMIT_MESSAGE
+        } else if (result.error.code === EXCEED_TOKEN_QUOTA_LIMIT) {
+          errMsg = EXCEED_TOKEN_QUOTA_LIMIT_MESSAGE
+        }
+        replyMsgData.content = errMsg
+      } else {
+        replyMsgData.content = result.content
+      }
     }
 
     console.log('replyMsgData:', replyMsgData)

cloudrunfunctions/cloudbase-agent/src/constant.ts

@@ -34,3 +34,41 @@ export const MSG_TYPE_TEXT = 'text'
 
 // 微信回调语音消息类型
 export const MSG_TYPE_VOICE = 'voice'
+
+// ==================== LLM 错误码 ====================
+
+/**
+ * 触发 HUNYUAN rateLimit
+ */
+export const HUNYUAN_RATE_LIMT_FAILED_CODE = 'HUNYUAN_RATE_LIMIT_FAILED'
+
+export const HUNYUAN_GENERATE_FAILED = 'HUNYUAN_GENERATE_FAILED'
+
+export const LLM_CONTENT_FILER = 'LLM_CONTENT_FILER'
+
+/**
+ * 请求聊天大模型失败
+ */
+export const REQUEST_LLM_ERROR_CODE = 'REQUEST_LLM_ERROR'
+
+export const REQUEST_LLM_ERROR_MESSAGE = '调用大模型失败，请稍后重试'
+
+/**
+ * LLM TOKEN 超限
+ */
+export const REQUEST_LLM_TOKEN_COUNT_EXCEED = 'REQUEST_LLM_TOKEN_COUNT_EXCEED'
+
+/**
+ * 大模型请求超限
+ */
+export const EXCEED_CONCURRENT_REQUEST_LIMIT = 'EXCEED_CONCURRENT_REQUEST_LIMIT'
+
+export const EXCEED_CONCURRENT_REQUEST_LIMIT_MESSAGE = '请求大模型并发超限，请稍后重试'
+
+/**
+ * 大模型 token 额度超限
+ */
+export const EXCEED_TOKEN_QUOTA_LIMIT = 'EXCEED_TOKEN_QUOTA_LIMIT'
+
+export const EXCEED_TOKEN_QUOTA_LIMIT_MESSAGE =
+  '大模型 Token 已耗尽，请通知开发者前往云开发平台进行处理。https://tcb.cloud.tencent.com/dev#/ai'

cloudrunfunctions/cloudbase-agent/src/llm.ts

@@ -1,11 +1,19 @@
 import { createDeepSeek } from '@ai-sdk/deepseek'
 import { createOpenAI } from '@ai-sdk/openai'
+import { APICallError } from 'ai'
 import * as ai from 'ai'
 import OpenAI from 'openai'
 
 import { BotContext } from './bot_context'
+import {
+  EXCEED_TOKEN_QUOTA_LIMIT,
+  EXCEED_TOKEN_QUOTA_LIMIT_MESSAGE,
+  REQUEST_LLM_ERROR_CODE,
+  REQUEST_LLM_ERROR_MESSAGE
+} from './constant'
 import { McpClient } from './mcp'
 import { getAccessToken } from './tcb'
+import { extractWithLodash, safeJsonParse } from './utils'
 
 const DEEPSEEK_PREFIX = 'deepseek'
 
@@ -28,8 +36,9 @@ export interface IMsgResult {
   content: string;
   finish_reason?: string;
   error?: {
-    name: string;
+    name?: string;
     message: string;
+    code?: string;
   };
   tool_call?: string;
   usage: object;
@@ -173,19 +182,31 @@ export class LLMCommunicator {
     messages: ChatCompletionMessage[],
     cb: (streamPart: ai.TextStreamPart<ai.ToolSet>) => void
   ) {
-    const { fullStream } = ai.streamText({
-      model: this.model,
-      tools: await this.mcpClient?.tools(),
-      maxSteps: 10,
-      messages: this.tarnsMessage([...messages]),
-      abortSignal: this.controller.signal,
-      onFinish: () => {
-        this.mcpClient?.close()
+    try {
+      const { fullStream } = ai.streamText({
+        model: this.model,
+        tools: await this.mcpClient?.tools(),
+        maxSteps: 10,
+        messages: this.tarnsMessage([...messages]),
+        abortSignal: this.controller.signal,
+        onFinish: () => {
+          this.mcpClient?.close()
+        }
+      })
+
+      for await (const streamPart of fullStream) {
+        cb(streamPart)
       }
-    })
+    } catch (e) {
+      cb?.({
+        type: 'error',
+        content: '',
+        step: 'error',
+        error: e,
+        finishReason: 'error'
+      })
 
-    for await (const streamPart of fullStream) {
-      cb(streamPart)
+      console.log('streamText_error:', JSON.stringify({ error: e, msg: e.message }))
     }
   }
 
@@ -296,15 +317,20 @@ export class LLMCommunicator {
             )
           } else if (streamPart.type === 'error') {
             // 对话异常
+            const gwLLMError = this.handlerAICallError(
+              (streamPart as any).error?.lastError || (streamPart as any).error
+            )
             result = {
               ...result,
               finish_reason: 'error',
-              error: {
-                name: 'LLMError',
-                message: streamPart.error as string
-              }
+              error: gwLLMError
             }
-            error = streamPart.error
+            error = (streamPart as any).error
+
+            console.log('stream_error:', JSON.stringify({
+              msg: (streamPart as any).error,
+              errMsg: (streamPart as any).error?.message
+            }))
 
             callMsg.push(result)
             this.botContext.bot.sseSender.send(
@@ -316,6 +342,9 @@ export class LLMCommunicator {
         }
       )
     } catch (error) {
+      const gwLLMError = this.handlerAICallError(
+        (error as any)?.lastError || error
+      )
       let result: IMsgResult = {
         type: 'error',
         created: Date.now(),
@@ -327,12 +356,10 @@ export class LLMCommunicator {
       result = {
         ...result,
         finish_reason: 'error',
-        error: {
-          name: 'LLMError',
-          message: error as string
-        }
+        error: gwLLMError
       }
-      // error = streamPart.error
+
+      console.log('stream_error:', JSON.stringify({ error, msg: (error as any)?.message }))
 
       callMsg.push(result)
       this.botContext.bot.sseSender.send(`data: ${JSON.stringify(result)}\n\n`)
@@ -369,8 +396,30 @@ export class LLMCommunicator {
       const generateTextRes = await ai.generateText(data)
       return cb(generateTextRes)
     } catch (error) {
-      console.log(error)
-      return {}
+      console.log('generateText_error:', JSON.stringify({ error, msg: (error as any)?.message }))
+      const gwLLMError = this.handlerAICallError(
+        (error as any)?.lastError || error
+      )
+      return { error: gwLLMError }
+    }
+  }
+
+  /**
+   * 解析 LLM 调用错误，转换为统一的内部错误码和友好消息
+   */
+  handlerAICallError (callError: APICallError) {
+    const responseBody = safeJsonParse(callError?.responseBody)
+    const result = extractWithLodash(responseBody, ['message', 'code'])
+    const code = result?.code?.includes(EXCEED_TOKEN_QUOTA_LIMIT)
+      ? EXCEED_TOKEN_QUOTA_LIMIT
+      : result?.code?.[0] || REQUEST_LLM_ERROR_CODE
+    let message = result?.message?.[0] || REQUEST_LLM_ERROR_MESSAGE
+    if (code === EXCEED_TOKEN_QUOTA_LIMIT) {
+      message = EXCEED_TOKEN_QUOTA_LIMIT_MESSAGE
+    }
+    return {
+      code,
+      message
     }
   }
 }

cloudrunfunctions/cloudbase-agent/src/utils.ts

@@ -1,5 +1,6 @@
 import * as crypto from 'crypto'
 import { customAlphabet } from 'nanoid'
+import lodash from 'lodash'
 
 export function genRandomStr (length: number): string {
   return crypto
@@ -23,3 +24,48 @@ export function safeJsonParse (jsonString: string, defaultValue = null) {
 export function randomId (len = 16) {
   return customAlphabet('0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz', len)()
 }
+
+// 解析嵌套对象，提取指定key的value
+export function extractWithLodash (obj: any, keys: string[], depth: number = 5): any {
+  if (depth <= 0) {
+    return {}
+  }
+
+  return lodash.reduce(
+    obj,
+    (result: any, value: any, key: string) => {
+      // 处理直接匹配的键
+      if (keys.includes(key) && !(lodash.isObject(value) || lodash.isArray(value))) {
+        result[key] = result[key] || []
+        result[key].push(value)
+      }
+
+      // 处理嵌套对象或数组
+      processNestedValue(result, value, keys, depth)
+
+      return result
+    },
+    {}
+  )
+}
+
+// 处理嵌套对象或数组的递归逻辑
+function processNestedValue (result: any, value: any, keys: string[], depth: number) {
+  if (lodash.isObject(value) && !lodash.isArray(value)) {
+    const nested = extractWithLodash(value, keys, depth - 1)
+    lodash.forEach(nested, (values, nestedKey) => {
+      result[nestedKey] = (result[nestedKey] || []).concat(values)
+    })
+  }
+
+  if (lodash.isArray(value)) {
+    value.forEach((item) => {
+      if (lodash.isObject(item)) {
+        const nested = extractWithLodash(item, keys, depth - 1)
+        lodash.forEach(nested, (values, nestedKey) => {
+          result[nestedKey] = (result[nestedKey] || []).concat(values)
+        })
+      }
+    })
+  }
+}