hyf-backend/th_agenter/services/llm_service.py

"""LLM service for workflow execution."""

import asyncio
from typing import List, Dict, Any, Optional, AsyncGenerator
from langchain_openai import ChatOpenAI
from langchain_core.messages import HumanMessage, AIMessage, SystemMessage

from ..models.llm_config import LLMConfig
from loguru import logger

class LLMService:
    """LLM服务，用于工作流中的大模型调用"""
    
    def __init__(self):
        pass
    
    async def chat_completion(
        self,
        model_config: LLMConfig,
        messages: List[Dict[str, str]],
        temperature: Optional[float] = None,
        max_tokens: Optional[int] = None
    ) -> str:
        """调用大模型进行对话完成"""
        # 处理 base_url：如果包含 /chat/completions，需要移除
        base_url = model_config.base_url
        if base_url and '/chat/completions' in base_url:
            # 移除 /chat/completions 后缀，ChatOpenAI 会自动添加
            base_url = base_url.replace('/chat/completions', '').rstrip('/')
            logger.debug(f"调整 base_url: {model_config.base_url} -> {base_url}")
        
        # 处理 SiliconFlow 的模型名称格式
        model_name = model_config.model_name
        if 'siliconflow' in (base_url or '').lower() and '/' not in model_name:
            # SiliconFlow 需要 org/model 格式，尝试自动转换
            model_name_lower = model_name.lower()
            if 'deepseek' in model_name_lower or 'r1' in model_name_lower:
                # 尝试常见的 DeepSeek 模型格式
                if 'r1' in model_name_lower:
                    model_name = 'deepseek-ai/DeepSeek-R1'
                elif 'v3' in model_name_lower:
                    model_name = 'deepseek-ai/DeepSeek-V3'
                else:
                    model_name = f'deepseek-ai/{model_name}'
                logger.debug(f"调整 SiliconFlow 模型名称: {model_config.model_name} -> {model_name}")
        
        try:
            # 创建LangChain ChatOpenAI实例
            llm = ChatOpenAI(
                model=model_name,
                api_key=model_config.api_key,
                base_url=base_url,
                temperature=temperature or model_config.temperature,
                max_tokens=max_tokens or model_config.max_tokens,
                streaming=False
            )
            
            # 转换消息格式
            langchain_messages = []
            for msg in messages:
                role = msg.get("role", "user")
                content = msg.get("content", "")
                
                if role == "system":
                    langchain_messages.append(SystemMessage(content=content))
                elif role == "user":
                    langchain_messages.append(HumanMessage(content=content))
                elif role == "assistant":
                    langchain_messages.append(AIMessage(content=content))
            
            # 调用LLM
            response = await llm.ainvoke(langchain_messages)
            
            # 返回响应内容
            return response.content
            
        except Exception as e:
            # 提取详细的错误信息
            error_detail = str(e)
            
            # 尝试从异常对象中提取更多信息
            if hasattr(e, 'response'):
                # HTTP 响应错误
                if hasattr(e.response, 'status_code'):
                    error_detail = f"HTTP {e.response.status_code}: {error_detail}"
                if hasattr(e.response, 'text'):
                    try:
                        import json
                        error_body = json.loads(e.response.text)
                        if isinstance(error_body, dict):
                            if 'message' in error_body:
                                error_detail = f"{error_detail} - {error_body['message']}"
                            if 'error' in error_body:
                                error_info = error_body['error']
                                if isinstance(error_info, dict) and 'message' in error_info:
                                    error_detail = f"{error_detail} - {error_info['message']}"
                    except:
                        pass
            
            # 添加模型配置信息到错误消息中（使用处理后的配置）
            model_info = f"模型: {model_name}, base_url: {base_url}"
            if 'Not Found' in error_detail or '404' in error_detail:
                error_detail = f"{error_detail} ({model_info})。可能的原因：1) 模型名称格式不正确（SiliconFlow需要org/model格式，如deepseek-ai/DeepSeek-R1）；2) base_url配置错误；3) API端点不存在"
            elif '403' in error_detail or 'account balance' in error_detail.lower() or 'insufficient' in error_detail.lower():
                error_detail = f"{error_detail} ({model_info})。可能的原因：账户余额不足或API密钥权限不足"
            elif '401' in error_detail or 'authentication' in error_detail.lower():
                error_detail = f"{error_detail} ({model_info})。可能的原因：API密钥无效或已过期"
            else:
                error_detail = f"{error_detail} ({model_info})"
            
            logger.error(f"LLM调用失败: {error_detail}")
            raise Exception(f"LLM调用失败: {error_detail}")
    
    async def chat_completion_stream(
        self,
        model_config: LLMConfig,
        messages: List[Dict[str, str]],
        temperature: Optional[float] = None,
        max_tokens: Optional[int] = None
    ) -> AsyncGenerator[str, None]:
        """调用大模型进行流式对话完成"""
        try:
            # 创建LangChain ChatOpenAI实例（流式）
            llm = ChatOpenAI(
                model=model_config.model_name,
                api_key=model_config.api_key,
                base_url=model_config.base_url,
                temperature=temperature or model_config.temperature,
                max_tokens=max_tokens or model_config.max_tokens,
                streaming=True
            )
            
            # 转换消息格式
            langchain_messages = []
            for msg in messages:
                role = msg.get("role", "user")
                content = msg.get("content", "")
                
                if role == "system":
                    langchain_messages.append(SystemMessage(content=content))
                elif role == "user":
                    langchain_messages.append(HumanMessage(content=content))
                elif role == "assistant":
                    langchain_messages.append(AIMessage(content=content))
            
            # 流式调用LLM
            async for chunk in llm.astream(langchain_messages):
                if hasattr(chunk, 'content') and chunk.content:
                    yield chunk.content
            
        except Exception as e:
            logger.error(f"LLM流式调用失败: {str(e)}")
            raise Exception(f"LLM流式调用失败: {str(e)}")
    
    def get_model_info(self, model_config: LLMConfig) -> Dict[str, Any]:
        """获取模型信息"""
        return {
            "id": model_config.id,
            "name": model_config.model_name,
            "provider": model_config.provider,
            "base_url": model_config.base_url,
            "temperature": model_config.temperature,
            "max_tokens": model_config.max_tokens,
            "is_active": model_config.is_active
        }