""" LLM适配器 - 使用DeepSeek LLM API检测品牌引用 """ import asyncio import json import logging import random import re from typing import Optional from app.schemas.scoring import CitationResult from app.config import settings logger = logging.getLogger(__name__) BRAND_CITATION_PROMPT = """分析以下AI搜索查询中是否提到了目标品牌。 查询关键词: {keyword} 目标品牌: {brand_name} 品牌别名: {brand_aliases} 返回JSON格式: {{"cited": true/false, "position": 1, "citation_text": "...", "sentiment": "positive/neutral/negative", "confidence": 0.95}} """ class LLMAdapterError(Exception): """LLM适配器异常""" pass class LLMAdapter: """LLM适配器 - 使用 OpenAI 兼容协议检测品牌引用（支持百炼/DashScope/DeepSeek）""" def __init__(self, api_key: Optional[str] = None, max_retries: int = 3): """ 初始化LLM适配器 Args: api_key: API密钥，默认优先使用 OPENAI_API_KEY（百炼/DashScope），其次 DEEPSEEK_API_KEY max_retries: 最大重试次数 """ self.api_key = ( api_key or getattr(settings, 'OPENAI_API_KEY', None) or getattr(settings, 'DEEPSEEK_API_KEY', None) ) # base_url 优先 OPENAI_BASE_URL，其次 DEEPSEEK_BASE_URL self.base_url = ( getattr(settings, 'OPENAI_BASE_URL', None) or getattr(settings, 'DEEPSEEK_BASE_URL', 'https://api.deepseek.com/v1') ) # model 优先 OPENAI_MODEL，其次 DEFAULT_LLM_MODEL self.model = ( getattr(settings, 'OPENAI_MODEL', None) or getattr(settings, 'DEFAULT_LLM_MODEL', 'qwen3-coder-plus') or 'qwen3-coder-plus' ) self.max_retries = max_retries self._client = None @property def client(self): """延迟初始化 OpenAI 兼容客户端""" if self._client is None: try: from openai import OpenAI self._client = OpenAI( api_key=self.api_key, base_url=self.base_url, ) except ImportError: raise LLMAdapterError("请安装openai库: pip install openai") return self._client def _build_prompt(self, keyword: str, brand_name: str, brand_aliases: list[str]) -> str: """构建Prompt""" aliases_str = ", ".join(brand_aliases) if brand_aliases else "无" return BRAND_CITATION_PROMPT.format( keyword=keyword, brand_name=brand_name, brand_aliases=aliases_str ) async def query_brand_citation( self, keyword: str, brand_name: str, brand_aliases: list[str] ) -> CitationResult: """ 使用LLM检测品牌引用 Args: keyword: 查询关键词 brand_name: 目标品牌名称 brand_aliases: 品牌别名列表 Returns: CitationResult: 包含cited, position, citation_text, sentiment, confidence Raises: LLMAdapterError: API调用或解析失败 """ if not settings.ENABLE_LLM: logger.info("LLM调用已禁用 (ENABLE_LLM=False)，返回模拟数据") return self._get_mock_result(keyword, brand_name, brand_aliases) prompt = self._build_prompt(keyword, brand_name, brand_aliases) last_error = None for attempt in range(self.max_retries): try: response = await self._call_deepseek(prompt) return self._parse_response(response) except Exception as e: last_error = e logger.warning( f"LLM API调用失败 (尝试 {attempt + 1}/{self.max_retries}): {e}" ) raise LLMAdapterError(f"LLM API调用失败，已重试{self.max_retries}次: {last_error}") def _get_mock_result( self, keyword: str, brand_name: str, brand_aliases: list[str] ) -> CitationResult: """ 生成模拟结果（当LLM禁用时使用） 随机决定是否引用，模拟真实场景的数据分布 """ cited = random.random() < 0.6 sentiment_options = ["positive", "neutral", "negative"] sentiment = random.choice(sentiment_options) if cited: position = random.randint(1, 10) citation_text = f'模拟引用：在搜索"{keyword}"时，提到了{brand_name}品牌及其相关产品。' else: position = None citation_text = "" return CitationResult( cited=cited, position=position, citation_text=citation_text, sentiment=sentiment, confidence=round(random.uniform(0.7, 0.99), 2) ) async def _call_deepseek(self, prompt: str) -> dict: """ 调用DeepSeek API Args: prompt: 提示词 Returns: API响应的JSON解析结果 Raises: LLMAdapterError: API调用失败 """ try: # 在线程池中执行同步的API调用 response = await asyncio.to_thread( self._sync_call_deepseek, prompt ) return response except json.JSONDecodeError as e: raise LLMAdapterError(f"JSON解析失败: {e}") except Exception as e: raise LLMAdapterError(f"API调用失败: {e}") def _sync_call_deepseek(self, prompt: str) -> dict: """ 同步调用DeepSeek API（在线程池中执行） Args: prompt: 提示词 Returns: API响应的JSON解析结果 """ response = self.client.chat.completions.create( model=self.model, messages=[ { "role": "user", "content": prompt } ], temperature=0.1, max_tokens=500, ) content = response.choices[0].message.content if not content: raise LLMAdapterError("API返回空响应") # 提取JSON（可能包裹在```json block中） json_str = self._extract_json(content) return json.loads(json_str) def _extract_json(self, text: str) -> str: """从文本中提取JSON""" # 尝试直接解析 try: json.loads(text) return text except json.JSONDecodeError: pass # 尝试从代码块中提取 json_pattern = r'```(?:json)?\s*([\s\S]*?)\s*```' match = re.search(json_pattern, text) if match: return match.group(1).strip() # 尝试找到第一个{到最后一个}之间的内容 first_brace = text.find('{') last_brace = text.rfind('}') if first_brace != -1 and last_brace != -1 and last_brace > first_brace: return text[first_brace:last_brace + 1] raise LLMAdapterError(f"无法从响应中提取JSON: {text[:200]}") def _parse_response(self, response: dict) -> CitationResult: """ 解析API响应 Args: response: API返回的字典 Returns: CitationResult对象 Raises: LLMAdapterError: 解析失败 """ try: required_fields = ['cited', 'sentiment', 'confidence'] for field in required_fields: if field not in response: raise LLMAdapterError(f"响应缺少必需字段: {field}") cited = bool(response['cited']) sentiment = str(response.get('sentiment', 'neutral')).lower() if sentiment not in ['positive', 'neutral', 'negative']: sentiment = 'neutral' # 验证position position = response.get('position') if position is not None: position = int(position) if position < 1: position = None # 验证confidence confidence = float(response.get('confidence', 0.5)) confidence = max(0.0, min(1.0, confidence)) citation_text = response.get('citation_text') if citation_text and len(citation_text) > 500: citation_text = citation_text[:500] return CitationResult( cited=cited, position=position, citation_text=citation_text, sentiment=sentiment, confidence=confidence ) except (ValueError, TypeError) as e: raise LLMAdapterError(f"解析响应失败: {e}") async def close(self): """关闭客户端连接""" if self._client is not None: try: # OpenAI/DeepSeek客户端不需要显式关闭 pass except Exception as e: logger.warning(f"关闭LLM客户端时出错: {e}") finally: self._client = None