app/utils/llm_client.py

"""
LLM客户端封装
统一使用OpenAI格式调用
"""

import json
import re
from typing import Optional, Dict, Any, List
from openai import OpenAI

from ..config import Config
from .api_key_runtime import call_with_limit_rotation, get_llm_key


class LLMClient:
    """LLM客户端"""
    
    def __init__(
        self,
        api_key: Optional[str] = None,
        base_url: Optional[str] = None,
        model: Optional[str] = None
    ):
        self._fixed_api_key = (api_key or "").strip() or None
        self.base_url = base_url or Config.LLM_BASE_URL
        self.model = model or Config.LLM_MODEL_NAME

        if not self._fixed_api_key and not Config.get_llm_api_keys():
            raise ValueError("LLM API key not configured (set LLM_1 or LLM_API_KEY)")
    
    def chat(
        self,
        messages: List[Dict[str, str]],
        temperature: float = 0.7,
        max_tokens: int = 4096,
        response_format: Optional[Dict] = None
    ) -> str:
        """
        发送聊天请求
        
        Args:
            messages: 消息列表
            temperature: 温度参数
            max_tokens: 最大token数
            response_format: 响应格式（如JSON模式）
            
        Returns:
            模型响应文本
        """
        kwargs = {
            "model": self.model,
            "messages": messages,
            "temperature": temperature,
            "max_tokens": max_tokens,
        }
        
        if response_format:
            kwargs["response_format"] = response_format

        def _create():
            key = (self._fixed_api_key or get_llm_key() or "").strip()
            if not key:
                raise ValueError("LLM API key not configured (set LLM_1 / LLM_2 or LLM_API_KEY)")
            client = OpenAI(api_key=key, base_url=self.base_url)
            return client.chat.completions.create(**kwargs)

        response = (
            _create()
            if self._fixed_api_key
            else call_with_limit_rotation(_create)
        )
        content = response.choices[0].message.content
        # 部分模型（如MiniMax M2.5）会在content中包含<think>思考内容，需要移除
        content = re.sub(r'<think>[\s\S]*?</think>', '', content).strip()
        return content
    
    def chat_json(
        self,
        messages: List[Dict[str, str]],
        temperature: float = 0.3,
        max_tokens: int = 4096
    ) -> Dict[str, Any]:
        """
        发送聊天请求并返回JSON
        
        Args:
            messages: 消息列表
            temperature: 温度参数
            max_tokens: 最大token数
            
        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()

        try:
            return json.loads(cleaned_response)
        except json.JSONDecodeError:
            raise ValueError(f"LLM返回的JSON格式无效: {cleaned_response}")