From 52c177fd662c805e43bf44910b6ad918591b0c1d Mon Sep 17 00:00:00 2001 From: LoryGlory Date: Mon, 23 Mar 2026 15:45:09 +0100 Subject: [PATCH] feat(llm): add structured output reliability to LLMClient and refactor services - chat_json() gains configurable retry (max_attempts), temperature backoff (temperature_step), optional retry delay, and a fallback_parser hook for service-specific rescue logic - _clean_response_text: strip tags and markdown code fences - _fix_truncated_json: use unescaped-quote parity instead of last-char check to avoid spuriously quoting numeric values; fixes broken repair for arrays - _try_fix_json: generic near-valid JSON salvage (newline normalisation, control-character stripping, greedy object extraction) - simulation_config_generator: replace raw OpenAI client with LLMClient, remove duplicated local retry loop and truncated-JSON repair, pass service-specific config salvage as fallback_parser - oasis_profile_generator: same refactor; keep rule-based profile fallback Co-Authored-By: Claude Sonnet 4.6 --- .../app/services/oasis_profile_generator.py | 96 +++++------- .../services/simulation_config_generator.py | 90 +++-------- backend/app/utils/llm_client.py | 142 +++++++++++++++--- 3 files changed, 170 insertions(+), 158 deletions(-) diff --git a/backend/app/services/oasis_profile_generator.py b/backend/app/services/oasis_profile_generator.py index 57836c53..53c6ad11 100644 --- a/backend/app/services/oasis_profile_generator.py +++ b/backend/app/services/oasis_profile_generator.py @@ -15,10 +15,10 @@ from typing import Dict, Any, List, Optional from dataclasses import dataclass, field from datetime import datetime -from openai import OpenAI from zep_cloud.client import Zep from ..config import Config +from ..utils.llm_client import LLMClient from ..utils.logger import get_logger from .zep_entity_reader import EntityNode, ZepEntityReader @@ -192,9 +192,10 @@ class OasisProfileGenerator: if not self.api_key: raise ValueError("LLM_API_KEY 未配置") - self.client = OpenAI( + self.llm_client = LLMClient( api_key=self.api_key, - base_url=self.base_url + base_url=self.base_url, + model=self.model_name ) # Zep客户端用于检索丰富上下文 @@ -520,64 +521,36 @@ class OasisProfileGenerator: entity_name, entity_type, entity_summary, entity_attributes, context ) - # 尝试多次生成，直到成功或达到最大重试次数 - max_attempts = 3 - last_error = None - - for attempt in range(max_attempts): - try: - response = self.client.chat.completions.create( - model=self.model_name, - messages=[ - {"role": "system", "content": self._get_system_prompt(is_individual)}, - {"role": "user", "content": prompt} - ], - response_format={"type": "json_object"}, - temperature=0.7 - (attempt * 0.1) # 每次重试降低温度 - # 不设置max_tokens，让LLM自由发挥 - ) - - content = response.choices[0].message.content - - # 检查是否被截断（finish_reason不是'stop'） - finish_reason = response.choices[0].finish_reason - if finish_reason == 'length': - logger.warning(f"LLM输出被截断 (attempt {attempt+1}), 尝试修复...") - content = self._fix_truncated_json(content) - - # 尝试解析JSON - try: - result = json.loads(content) - - # 验证必需字段 - if "bio" not in result or not result["bio"]: - result["bio"] = entity_summary[:200] if entity_summary else f"{entity_type}: {entity_name}" - if "persona" not in result or not result["persona"]: - result["persona"] = entity_summary or f"{entity_name}是一个{entity_type}。" - - return result - - except json.JSONDecodeError as je: - logger.warning(f"JSON解析失败 (attempt {attempt+1}): {str(je)[:80]}") - - # 尝试修复JSON - result = self._try_fix_json(content, entity_name, entity_type, entity_summary) - if result.get("_fixed"): - del result["_fixed"] - return result - - last_error = je - - except Exception as e: - logger.warning(f"LLM调用失败 (attempt {attempt+1}): {str(e)[:80]}") - last_error = e - import time - time.sleep(1 * (attempt + 1)) # 指数退避 - - logger.warning(f"LLM生成人设失败（{max_attempts}次尝试）: {last_error}, 使用规则生成") - return self._generate_profile_rule_based( - entity_name, entity_type, entity_summary, entity_attributes - ) + try: + result = self.llm_client.chat_json( + messages=[ + {"role": "system", "content": self._get_system_prompt(is_individual)}, + {"role": "user", "content": prompt} + ], + temperature=0.7, + max_tokens=None, + max_attempts=3, + temperature_step=0.1, + fallback_parser=lambda content: self._try_fix_json( + content, entity_name, entity_type, entity_summary + ), + retry_delay_seconds=1.0 + ) + + if "bio" not in result or not result["bio"]: + result["bio"] = entity_summary[:200] if entity_summary else f"{entity_type}: {entity_name}" + if "persona" not in result or not result["persona"]: + result["persona"] = entity_summary or f"{entity_name}是一个{entity_type}。" + + if result.get("_fixed"): + del result["_fixed"] + + return result + except Exception as e: + logger.warning(f"LLM生成人设失败（3次尝试）: {e}, 使用规则生成") + return self._generate_profile_rule_based( + entity_name, entity_type, entity_summary, entity_attributes + ) def _fix_truncated_json(self, content: str) -> str: """修复被截断的JSON（输出被max_tokens限制截断）""" @@ -1197,4 +1170,3 @@ class OasisProfileGenerator: """[已废弃] 请使用 save_profiles() 方法""" logger.warning("save_profiles_to_json已废弃，请使用save_profiles方法") self.save_profiles(profiles, file_path, platform) - diff --git a/backend/app/services/simulation_config_generator.py b/backend/app/services/simulation_config_generator.py index cc362508..0cb8ab1e 100644 --- a/backend/app/services/simulation_config_generator.py +++ b/backend/app/services/simulation_config_generator.py @@ -16,9 +16,8 @@ from typing import Dict, Any, List, Optional, Callable from dataclasses import dataclass, field, asdict from datetime import datetime -from openai import OpenAI - from ..config import Config +from ..utils.llm_client import LLMClient from ..utils.logger import get_logger from .zep_entity_reader import EntityNode, ZepEntityReader @@ -234,9 +233,10 @@ class SimulationConfigGenerator: if not self.api_key: raise ValueError("LLM_API_KEY 未配置") - self.client = OpenAI( + self.llm_client = LLMClient( api_key=self.api_key, - base_url=self.base_url + base_url=self.base_url, + model=self.model_name ) def generate_config( @@ -432,78 +432,23 @@ class SimulationConfigGenerator: def _call_llm_with_retry(self, prompt: str, system_prompt: str) -> Dict[str, Any]: """带重试的LLM调用，包含JSON修复逻辑""" - import re - - max_attempts = 3 - last_error = None - - for attempt in range(max_attempts): - try: - response = self.client.chat.completions.create( - model=self.model_name, - messages=[ - {"role": "system", "content": system_prompt}, - {"role": "user", "content": prompt} - ], - response_format={"type": "json_object"}, - temperature=0.7 - (attempt * 0.1) # 每次重试降低温度 - # 不设置max_tokens，让LLM自由发挥 - ) - - content = response.choices[0].message.content - finish_reason = response.choices[0].finish_reason - - # 检查是否被截断 - if finish_reason == 'length': - logger.warning(f"LLM输出被截断 (attempt {attempt+1})") - content = self._fix_truncated_json(content) - - # 尝试解析JSON - try: - return json.loads(content) - except json.JSONDecodeError as e: - logger.warning(f"JSON解析失败 (attempt {attempt+1}): {str(e)[:80]}") - - # 尝试修复JSON - fixed = self._try_fix_config_json(content) - if fixed: - return fixed - - last_error = e - - except Exception as e: - logger.warning(f"LLM调用失败 (attempt {attempt+1}): {str(e)[:80]}") - last_error = e - import time - time.sleep(2 * (attempt + 1)) - - raise last_error or Exception("LLM调用失败") - - def _fix_truncated_json(self, content: str) -> str: - """修复被截断的JSON""" - content = content.strip() - - # 计算未闭合的括号 - open_braces = content.count('{') - content.count('}') - open_brackets = content.count('[') - content.count(']') - - # 检查是否有未闭合的字符串 - if content and content[-1] not in '",}]': - content += '"' - - # 闭合括号 - content += ']' * open_brackets - content += '}' * open_braces - - return content + return self.llm_client.chat_json( + messages=[ + {"role": "system", "content": system_prompt}, + {"role": "user", "content": prompt} + ], + temperature=0.7, + max_tokens=None, + max_attempts=3, + temperature_step=0.1, + fallback_parser=self._try_fix_config_json, + retry_delay_seconds=2.0 + ) def _try_fix_config_json(self, content: str) -> Optional[Dict[str, Any]]: """尝试修复配置JSON""" import re - - # 修复被截断的情况 - content = self._fix_truncated_json(content) - + # 提取JSON部分 json_match = re.search(r'\{[\s\S]*\}', content) if json_match: @@ -984,4 +929,3 @@ class SimulationConfigGenerator: "influence_weight": 1.0 } - diff --git a/backend/app/utils/llm_client.py b/backend/app/utils/llm_client.py index 6c1a81f4..953327ab 100644 --- a/backend/app/utils/llm_client.py +++ b/backend/app/utils/llm_client.py @@ -4,11 +4,15 @@ LLM客户端封装 """ import json +import time import re -from typing import Optional, Dict, Any, List +from typing import Optional, Dict, Any, List, Callable from openai import OpenAI from ..config import Config +from .logger import get_logger + +logger = get_logger('mirofish.llm_client') class LLMClient: @@ -36,7 +40,7 @@ class LLMClient: self, messages: List[Dict[str, str]], temperature: float = 0.7, - max_tokens: int = 4096, + max_tokens: Optional[int] = 4096, response_format: Optional[Dict] = None ) -> str: """ @@ -55,23 +59,27 @@ class LLMClient: "model": self.model, "messages": messages, "temperature": temperature, - "max_tokens": max_tokens, } + + if max_tokens is not None: + kwargs["max_tokens"] = max_tokens if response_format: kwargs["response_format"] = response_format response = self.client.chat.completions.create(**kwargs) - content = response.choices[0].message.content - # 部分模型（如MiniMax M2.5）会在content中包含思考内容，需要移除 - content = re.sub(r'[\s\S]*?', '', content).strip() - return content - + content = response.choices[0].message.content or "" + return self._clean_response_text(content) + def chat_json( self, messages: List[Dict[str, str]], temperature: float = 0.3, - max_tokens: int = 4096 + max_tokens: Optional[int] = 4096, + max_attempts: int = 1, + temperature_step: float = 0.0, + fallback_parser: Optional[Callable[[str], Optional[Dict[str, Any]]]] = None, + retry_delay_seconds: float = 0.0 ) -> Dict[str, Any]: """ 发送聊天请求并返回JSON @@ -84,20 +92,108 @@ class LLMClient: Returns: 解析后的JSON对象 """ - response = self.chat( - messages=messages, - temperature=temperature, - max_tokens=max_tokens, - response_format={"type": "json_object"} - ) - # 清理markdown代码块标记 - cleaned_response = response.strip() - cleaned_response = re.sub(r'^```(?:json)?\s*\n?', '', cleaned_response, flags=re.IGNORECASE) - cleaned_response = re.sub(r'\n?```\s*$', '', cleaned_response) - cleaned_response = cleaned_response.strip() + last_error: Optional[Exception] = None + last_response = "" + + for attempt in range(max_attempts): + current_temperature = max(0.0, temperature - (attempt * temperature_step)) + + try: + kwargs = { + "model": self.model, + "messages": messages, + "temperature": current_temperature, + "response_format": {"type": "json_object"} + } + + if max_tokens is not None: + kwargs["max_tokens"] = max_tokens + + response = self.client.chat.completions.create(**kwargs) + raw_content = response.choices[0].message.content or "" + finish_reason = response.choices[0].finish_reason + + cleaned_response = self._clean_response_text(raw_content) + if finish_reason == 'length': + logger.warning(f"LLM输出被截断 (attempt {attempt + 1})") + cleaned_response = self._fix_truncated_json(cleaned_response) + + last_response = cleaned_response + + try: + return self._parse_json_response(cleaned_response) + except json.JSONDecodeError as parse_error: + logger.warning(f"JSON解析失败 (attempt {attempt + 1}): {str(parse_error)[:80]}") + + fixed = self._try_fix_json(cleaned_response) + if fixed is not None: + return fixed + + if fallback_parser is not None: + fallback_result = fallback_parser(cleaned_response) + if fallback_result is not None: + return fallback_result + + last_error = parse_error + + except Exception as exc: + logger.warning(f"LLM调用失败 (attempt {attempt + 1}): {str(exc)[:80]}") + last_error = exc + + if attempt < max_attempts - 1 and retry_delay_seconds > 0: + time.sleep(retry_delay_seconds * (attempt + 1)) + + raise ValueError(f"LLM返回的JSON格式无效: {last_response}") from last_error + + def _clean_response_text(self, content: str) -> str: + """清理模型响应中的思考内容和Markdown包裹。""" + cleaned = re.sub(r'[\s\S]*?', '', content).strip() + cleaned = re.sub(r'^```(?:json)?\s*\n?', '', cleaned, flags=re.IGNORECASE) + cleaned = re.sub(r'\n?```\s*$', '', cleaned) + return cleaned.strip() + + def _parse_json_response(self, content: str) -> Dict[str, Any]: + return json.loads(content) + + def _fix_truncated_json(self, content: str) -> str: + """修复被截断的JSON内容。""" + content = content.strip() + + # If the number of unescaped quotes is odd we are inside an open string. + unescaped_quote_count = len(re.findall(r'(? Optional[Dict[str, Any]]: + """尝试从近似JSON内容中恢复结构化对象。""" + content = self._fix_truncated_json(content) + json_match = re.search(r'\{[\s\S]*\}', content) + if not json_match: + return None + + json_str = json_match.group() + + def fix_string_newlines(match: re.Match[str]) -> str: + value = match.group(0) + value = value.replace('\n', ' ').replace('\r', ' ') + value = re.sub(r'\s+', ' ', value) + return value + + json_str = re.sub(r'"[^"\\]*(?:\\.[^"\\]*)*"', fix_string_newlines, json_str) try: - return json.loads(cleaned_response) + return json.loads(json_str) except json.JSONDecodeError: - raise ValueError(f"LLM返回的JSON格式无效: {cleaned_response}") - + json_str = re.sub(r'[\x00-\x1f\x7f-\x9f]', ' ', json_str) + json_str = re.sub(r'\s+', ' ', json_str) + try: + return json.loads(json_str) + except json.JSONDecodeError: + return None