feat: 统一LLM错误处理并添加系统错误回复配置

在多个LLM提供者中移除try-catch块，将错误处理统一到connection.py的流处理层
添加system_error_response配置项，支持自定义系统错误时的回复内容
在意图识别和流处理中捕获异常时返回配置的错误回复，避免硬编码错误信息

Fixes #2075

main/xiaozhi-server/config.yaml

@@ -202,6 +202,9 @@ prompt: |
 # 默认系统提示词模板文件
 prompt_template: agent-base-prompt.txt
 
+# 系统错误时的回复
+system_error_response: "主人，小智现在有点忙，我们稍后再试吧。"
+
 # 结束语prompt
 end_prompt:
   enable: true # 是否开启结束语

main/xiaozhi-server/core/connection.py

@@ -40,8 +40,10 @@ from config.logger import setup_logging, build_module_string, create_connection_
 from config.manage_api_client import DeviceNotFoundException, DeviceBindException
 from core.utils.prompt_manager import PromptManager
 from core.utils.voiceprint_provider import VoiceprintProvider
+from core.utils.util import get_system_error_response
 from core.utils import textUtils
 
+
 TAG = __name__
 
 auto_import_modules("plugins_func.functions")
@@ -869,6 +871,7 @@ class ConnectionHandler:
         content_arguments = ""
         self.client_abort = False
         emotion_flag = True
+        try:
             for response in llm_responses:
                 if self.client_abort:
                     break
@@ -909,6 +912,26 @@ class ConnectionHandler:
                                 content_detail=content,
                             )
                         )
+        except Exception as e:
+            self.logger.bind(tag=TAG).error(f"LLM stream processing error: {e}")
+            self.tts.tts_text_queue.put(
+                TTSMessageDTO(
+                    sentence_id=self.sentence_id,
+                    sentence_type=SentenceType.LAST,
+                    content_type=ContentType.TEXT,
+                    content_detail=get_system_error_response(self.config),
+                )
+            )
+            if depth == 0:
+                self.tts.tts_text_queue.put(
+                    TTSMessageDTO(
+                        sentence_id=self.sentence_id,
+                        sentence_type=SentenceType.LAST,
+                        content_type=ContentType.ACTION,
+                    )
+                )
+                self.llm_finish_task = True
+            return
         # 处理function call
         if tool_call_flag:
             bHasError = False

main/xiaozhi-server/core/providers/intent/intent_llm/intent_llm.py

@@ -2,11 +2,14 @@ from typing import List, Dict
 from ..base import IntentProviderBase
 from plugins_func.functions.play_music import initialize_music_handler
 from config.logger import setup_logging
+from core.utils.util import get_system_error_response
 import re
 import json
 import hashlib
 import time
 
+
+
 TAG = __name__
 logger = setup_logging()
 
@@ -115,12 +118,16 @@ class IntentProvider(IntentProviderBase):
         return prompt
 
     def replyResult(self, text: str, original_text: str):
+        try:
             llm_result = self.llm.response_no_stream(
                 system_prompt=text,
                 user_prompt="请根据以上内容，像人类一样说话的口吻回复用户，要求简洁，请直接返回结果。用户现在说："
                 + original_text,
             )
             return llm_result
+        except Exception as e:
+            logger.bind(tag=TAG).error(f"Error in generating reply result: {e}")
+            return get_system_error_response(self.config)
 
     async def detect_intent(self, conn, dialogue_history: List[Dict], text: str) -> str:
         if not self.llm:
@@ -194,9 +201,13 @@ class IntentProvider(IntentProviderBase):
         llm_start_time = time.time()
         logger.bind(tag=TAG).debug(f"开始LLM意图识别调用, 模型: {model_info}")
 
+        try:
             intent = self.llm.response_no_stream(
                 system_prompt=prompt_music, user_prompt=user_prompt
             )
+        except Exception as e:
+            logger.bind(tag=TAG).error(f"Error in intent detection LLM call: {e}")
+            return '{"function_call": {"name": "continue_chat"}}'
 
         # 记录LLM调用完成时间
         llm_time = time.time() - llm_start_time

main/xiaozhi-server/core/providers/llm/AliBL/AliBL.py

@@ -21,7 +21,6 @@ class LLMProvider(LLMProviderBase):
         check_model_key("AliBLLLM", self.api_key)
 
     def response(self, session_id, dialogue):
-        try:
         # 处理dialogue
         if self.is_No_prompt:
             dialogue.pop(0)
@@ -95,10 +94,6 @@ class LLMProvider(LLMProviderBase):
                     if chunk:
                         yield chunk
 
-        except Exception as e:
-            logger.bind(tag=TAG).error(f"【阿里百练API服务】响应异常: {e}")
-            yield "【LLM服务响应异常】"
-
     def response_with_functions(self, session_id, dialogue, functions=None):
         # 阿里百练当前未支持原生的 function call。为保持兼容，这里回退到普通文本流式输出。
         # 上层会按 (content, tool_calls) 的形式消费，这里始终返回 (token, None)

main/xiaozhi-server/core/providers/llm/base.py

@@ -11,7 +11,6 @@ class LLMProviderBase(ABC):
         pass
 
     def response_no_stream(self, system_prompt, user_prompt, **kwargs):
-        try:
         # 构造对话格式
         dialogue = [
             {"role": "system", "content": system_prompt},
@@ -22,10 +21,6 @@ class LLMProviderBase(ABC):
             result += part
         return result
     
-        except Exception as e:
-            logger.bind(tag=TAG).error(f"Error in Ollama response generation: {e}")
-            return "【LLM服务响应异常】"
-    
     def response_with_functions(self, session_id, dialogue, functions=None):
         """
         Default implementation for function calling (streaming)

main/xiaozhi-server/core/providers/llm/dify/dify.py

@@ -20,7 +20,6 @@ class LLMProvider(LLMProviderBase):
             logger.bind(tag=TAG).error(model_key_msg)
 
     def response(self, session_id, dialogue, **kwargs):
-        try:
         # 取最后一条用户消息
         last_msg = next(m for m in reversed(dialogue) if m["role"] == "user")
         conversation_id = self.session_conversation_map.get(session_id)
@@ -87,10 +86,6 @@ class LLMProvider(LLMProviderBase):
                         ):
                             yield event["answer"]
 
-        except Exception as e:
-            logger.bind(tag=TAG).error(f"Error in response generation: {e}")
-            yield "【服务响应异常】"
-
     def response_with_functions(self, session_id, dialogue, functions=None):
         if len(dialogue) == 2 and functions is not None and len(functions) > 0:
             # 第一次调用llm， 取最后一条用户消息，附加tool提示词

main/xiaozhi-server/core/providers/llm/fastgpt/fastgpt.py

@@ -19,7 +19,6 @@ class LLMProvider(LLMProviderBase):
             logger.bind(tag=TAG).error(model_key_msg)
 
     def response(self, session_id, dialogue, **kwargs):
-        try:
         # 取最后一条用户消息
         last_msg = next(m for m in reversed(dialogue) if m["role"] == "user")
 
@@ -63,10 +62,6 @@ class LLMProvider(LLMProviderBase):
                     except Exception as e:
                         continue
 
-        except Exception as e:
-            logger.bind(tag=TAG).error(f"Error in response generation: {e}")
-            yield "【服务响应异常】"
-
     def response_with_functions(self, session_id, dialogue, functions=None):
         logger.bind(tag=TAG).error(
             f"fastgpt暂未实现完整的工具调用（function call），建议使用其他意图识别"

main/xiaozhi-server/core/providers/llm/homeassistant/homeassistant.py

@@ -15,7 +15,6 @@ class LLMProvider(LLMProviderBase):
         self.api_url = f"{self.base_url}/api/conversation/process"  # 拼接完整的 API URL
 
     def response(self, session_id, dialogue, **kwargs):
-        try:
         # home assistant语音助手自带意图，无需使用xiaozhi ai自带的，只需要把用户说的话传递给home assistant即可
 
         # 提取最后一个 role 为 'user' 的 content
@@ -60,11 +59,6 @@ class LLMProvider(LLMProviderBase):
         else:
             logger.bind(tag=TAG).warning("API 返回数据中没有 speech 内容")
 
-        except RequestException as e:
-            logger.bind(tag=TAG).error(f"HTTP 请求错误: {e}")
-        except Exception as e:
-            logger.bind(tag=TAG).error(f"生成响应时出错: {e}")
-
     def response_with_functions(self, session_id, dialogue, functions=None):
         logger.bind(tag=TAG).error(
             f"homeassistant不支持（function call），建议使用其他意图识别"

main/xiaozhi-server/core/providers/llm/ollama/ollama.py

@@ -25,7 +25,6 @@ class LLMProvider(LLMProviderBase):
         self.is_qwen3 = self.model_name and self.model_name.lower().startswith("qwen3")
 
     def response(self, session_id, dialogue, **kwargs):
-        try:
         # 如果是qwen3模型，在用户最后一条消息中添加/no_think指令
         if self.is_qwen3:
             # 复制对话列表，避免修改原始对话
@@ -89,12 +88,7 @@ class LLMProvider(LLMProviderBase):
             except Exception as e:
                 logger.bind(tag=TAG).error(f"Error processing chunk: {e}")
 
-        except Exception as e:
-            logger.bind(tag=TAG).error(f"Error in Ollama response generation: {e}")
-            yield "【Ollama服务响应异常】"
-
     def response_with_functions(self, session_id, dialogue, functions=None):
-        try:
         # 如果是qwen3模型，在用户最后一条消息中添加/no_think指令
         if self.is_qwen3:
             # 复制对话列表，避免修改原始对话
@@ -169,7 +163,3 @@ class LLMProvider(LLMProviderBase):
             except Exception as e:
                 logger.bind(tag=TAG).error(f"Error processing function chunk: {e}")
                 continue
-
-        except Exception as e:
-            logger.bind(tag=TAG).error(f"Error in Ollama function call: {e}")
-            yield f"【Ollama服务响应异常: {str(e)}】", None

main/xiaozhi-server/core/providers/llm/openai/openai.py

@@ -56,7 +56,6 @@ class LLMProvider(LLMProviderBase):
         return dialogue
 
     def response(self, session_id, dialogue, **kwargs):
-        try:
         dialogue = self.normalize_dialogue(dialogue)
 
         request_params = {
@@ -77,6 +76,7 @@ class LLMProvider(LLMProviderBase):
             if value is not None:
                 request_params[key] = value
 
+        # raise ValueError("model_name is required")
         responses = self.client.chat.completions.create(**request_params)
 
         is_active = True
@@ -96,11 +96,7 @@ class LLMProvider(LLMProviderBase):
                 if is_active:
                     yield content
 
-        except Exception as e:
-            logger.bind(tag=TAG).error(f"Error in response generation: {e}")
-
     def response_with_functions(self, session_id, dialogue, functions=None, **kwargs):
-        try:
         dialogue = self.normalize_dialogue(dialogue)
 
         request_params = {
@@ -136,7 +132,3 @@ class LLMProvider(LLMProviderBase):
                     f"输出 {getattr(usage_info, 'completion_tokens', '未知')}，"
                     f"共计 {getattr(usage_info, 'total_tokens', '未知')}"
                 )
-
-        except Exception as e:
-            logger.bind(tag=TAG).error(f"Error in function call streaming: {e}")
-            yield f"【OpenAI服务响应异常: {e}】", None

main/xiaozhi-server/core/providers/llm/xinference/xinference.py

@@ -31,7 +31,6 @@ class LLMProvider(LLMProviderBase):
             raise
 
     def response(self, session_id, dialogue, **kwargs):
-        try:
         logger.bind(tag=TAG).debug(
             f"Sending request to Xinference with model: {self.model_name}, dialogue length: {len(dialogue)}"
         )
@@ -59,12 +58,7 @@ class LLMProvider(LLMProviderBase):
             except Exception as e:
                 logger.bind(tag=TAG).error(f"Error processing chunk: {e}")
 
-        except Exception as e:
-            logger.bind(tag=TAG).error(f"Error in Xinference response generation: {e}")
-            yield "【Xinference服务响应异常】"
-
     def response_with_functions(self, session_id, dialogue, functions=None):
-        try:
         logger.bind(tag=TAG).debug(
             f"Sending function call request to Xinference with model: {self.model_name}, dialogue length: {len(dialogue)}"
         )
@@ -89,10 +83,3 @@ class LLMProvider(LLMProviderBase):
                 yield content, tool_calls
             elif tool_calls:
                 yield None, tool_calls
-
-        except Exception as e:
-            logger.bind(tag=TAG).error(f"Error in Xinference function call: {e}")
-            yield {
-                "type": "content",
-                "content": f"【Xinference服务响应异常: {str(e)}】",
-            }

main/xiaozhi-server/core/providers/memory/mem_local_short/mem_local_short.py

@@ -162,6 +162,7 @@ class MemoryProvider(MemoryProviderBase):
         msgStr += f"当前时间：{time_str}"
 
         if self.save_to_file:
+            try:
                 result = self.llm.response_no_stream(
                     short_term_memory_prompt,
                     msgStr,
@@ -169,12 +170,11 @@ class MemoryProvider(MemoryProviderBase):
                     temperature=0.2,
                 )
                 json_str = extract_json_data(result)
-            try:
                 json.loads(json_str)  # 检查json格式是否正确
                 self.short_memory = json_str
                 self.save_memory_to_file()
             except Exception as e:
-                print("Error:", e)
+                logger.bind(tag=TAG).error(f"Error in saving memory: {e}")
         else:
             # 当save_to_file为False时，调用Java端的聊天记录总结接口
             summary_id = session_id if session_id else self.role_id

main/xiaozhi-server/core/utils/util.py

@@ -573,3 +573,14 @@ def validate_mcp_endpoint(mcp_endpoint: str) -> bool:
         return False
 
     return True
+
+def get_system_error_response(config: dict) -> str:
+    """获取系统错误时的回复
+
+    Args:
+        config: 配置字典
+
+    Returns:
+        str: 系统错误时的回复
+    """
+    return config.get("system_error_response", "主人，小智现在有点忙，我们稍后再试吧。")