cherry-ai/backend/app/services/chat_service.py

# File: backend/app/services/chat_service.py (Update with DB for history)
# Description: 封装 LangChain 聊天逻辑 (使用数据库存储和检索消息)

from langchain_openai import ChatOpenAI
from langchain_google_genai import ChatGoogleGenerativeAI
from langchain_core.prompts import ChatPromptTemplate, MessagesPlaceholder
from langchain_core.output_parsers import StrOutputParser
from langchain_core.messages import HumanMessage, AIMessage, BaseMessage, SystemMessage
from typing import Dict, List, Optional, Tuple
from sqlalchemy.ext.asyncio import AsyncSession
from sqlalchemy.future import select
from app.db.models import MessageModel, AssistantModel # Import DB models
from app.services.assistant_service import AssistantService # Use class directly
from app.models.pydantic_models import AssistantRead

class ChatService:
    """处理 AI 聊天交互的服务 (使用数据库历史)"""

    def __init__(self, default_api_key: str):
        self.default_api_key = default_api_key
        self.assistant_service = AssistantService() # Instantiate assistant service

    def _get_llm(self, assistant: AssistantRead) -> ChatOpenAI:
        # ... (remains the same) ...
        if assistant.model.startswith("gpt"):
             return ChatOpenAI(model=assistant.model, api_key=self.default_api_key, temperature=assistant.temperature)
        elif assistant.model.startswith("gemini"):
            return ChatGoogleGenerativeAI(
                model=assistant.model,
                api_key=self.default_api_key, # 或从助手配置中读取特定 key
                temperature=assistant.temperature
            )
        else:
             print(f"警告: 模型 {assistant.model} 未明确支持，尝试使用 ChatOpenAI")
             return ChatOpenAI(model=assistant.model, api_key=self.default_api_key, temperature=assistant.temperature)


    async def _get_chat_history(self, db: AsyncSession, session_id: str, limit: int = 10) -> Tuple[List[BaseMessage], int]:
        """从数据库加载指定会话的历史消息 (按 order 排序)"""
        stmt = (
            select(MessageModel)
            .filter(MessageModel.session_id == session_id)
            .order_by(MessageModel.order.desc()) # Get latest first
            .limit(limit)
        )
        result = await db.execute(stmt)
        db_messages = result.scalars().all()

        # Convert to LangChain messages (in correct order: oldest to newest)
        history: List[BaseMessage] = []
        max_order = 0
        for msg in reversed(db_messages): # Reverse to get chronological order
            if msg.sender == 'user':
                history.append(HumanMessage(content=msg.text))
            elif msg.sender == 'ai':
                history.append(AIMessage(content=msg.text))
            # Add handling for 'system' if needed
            max_order = max(max_order, msg.order) # Keep track of the latest order number

        return history, max_order

    async def _save_message(self, db: AsyncSession, session_id: str, sender: str, text: str, order: int):
        """将消息保存到数据库"""
        db_message = MessageModel(
            session_id=session_id,
            sender=sender,
            text=text,
            order=order
        )
        db.add(db_message)
        await db.flush() # Ensure it's added before potential commit
        print(f"消息已保存 (DB): Session={session_id}, Order={order}, Sender={sender}")


    async def get_ai_reply(self, db: AsyncSession, user_message: str, session_id: str, assistant_id: str) -> str:
        """获取 AI 回复，并保存用户消息和 AI 回复到数据库"""
        # 1. 获取助手配置
        assistant = await self.assistant_service.get_assistant(db, assistant_id)
        if not assistant:
            raise ValueError(f"找不到助手 ID: {assistant_id}")

        # 2. 获取历史记录和下一个序号
        current_chat_history, last_order = await self._get_chat_history(db, session_id)
        user_message_order = last_order + 1
        ai_message_order = last_order + 2

        # 3. 构建 Prompt
        prompt = ChatPromptTemplate.from_messages([
            SystemMessage(content=assistant.system_prompt),
            MessagesPlaceholder(variable_name="chat_history"),
            HumanMessage(content=user_message),
        ])

        # 4. 获取 LLM
        llm = self._get_llm(assistant)
        output_parser = StrOutputParser()
        chain = prompt | llm | output_parser

        try:
            # --- Save user message BEFORE calling LLM ---
            await self._save_message(db, session_id, 'user', user_message, user_message_order)

            # 5. 调用链获取回复
            ai_response = await chain.ainvoke({
                "input": user_message,
                "chat_history": current_chat_history, # Pass history fetched from DB
            })

            # --- Save AI response AFTER getting it ---
            await self._save_message(db, session_id, 'ai', ai_response, ai_message_order)

            # Note: We don't need to manage history in memory anymore (chat_history_db removed)
            return ai_response

        except Exception as e:
            # Consider rolling back the user message save if LLM call fails,
            # although often it's better to keep the user message.
            # await db.rollback() # Handled by get_db_session dependency on error
            print(f"调用 LangChain 时出错 (助手: {assistant_id}, 会话: {session_id}): {e}")
            raise Exception(f"AI 服务暂时不可用: {e}")

    async def generate_text(self, prompt_text: str, model_name: str = "gemini-2.0-flash", temperature: float = 0.5) -> str:
        # ... (remains the same) ...
        try:
            temp_llm = ChatGoogleGenerativeAI(
                model=model_name,
                api_key=self.default_api_key, # 或从助手配置中读取特定 key
                temperature=temperature
            )
            response = await temp_llm.ainvoke(prompt_text)
            return response.content
        except Exception as e:
            print(f"生成文本时出错: {e}")
            return "无法生成标题"


# ChatService instance is now created where needed or injected, no global instance here.