451 lines
17 KiB
Python
451 lines
17 KiB
Python
"""
|
||
多模型LLM服务 - 支持智谱AI和DeepSeek
|
||
"""
|
||
from typing import Optional, List, Dict, Any
|
||
from datetime import datetime, timedelta
|
||
from app.config import get_settings
|
||
from app.utils.logger import logger
|
||
|
||
# 智谱AI
|
||
try:
|
||
from zhipuai import ZhipuAI
|
||
ZHIPUAI_AVAILABLE = True
|
||
except ImportError:
|
||
ZHIPUAI_AVAILABLE = False
|
||
logger.warning("zhipuai包未安装")
|
||
|
||
# DeepSeek (使用OpenAI兼容接口)
|
||
try:
|
||
from openai import OpenAI
|
||
from openai import APIStatusError, APIError
|
||
OPENAI_AVAILABLE = True
|
||
except ImportError:
|
||
OPENAI_AVAILABLE = False
|
||
logger.warning("openai包未安装")
|
||
|
||
|
||
class MultiLLMService:
|
||
"""多模型LLM服务类"""
|
||
|
||
# 余额错误通知冷却时间（秒）
|
||
BALANCE_ERROR_COOLDOWN = 3600 # 1小时内只通知一次
|
||
|
||
def __init__(self):
|
||
"""初始化多模型LLM服务"""
|
||
settings = get_settings()
|
||
|
||
self.clients = {}
|
||
self.current_model = None
|
||
self.model_info = {}
|
||
|
||
# 余额错误通知时间记录
|
||
self._balance_error_notified = {} # {provider: last_notified_time}
|
||
|
||
# 初始化智谱AI
|
||
if ZHIPUAI_AVAILABLE and settings.zhipuai_api_key:
|
||
try:
|
||
api_key = settings.zhipuai_api_key.strip()
|
||
if '.' in api_key and len(api_key) > 10:
|
||
self.clients['zhipu'] = ZhipuAI(api_key=api_key)
|
||
self.model_info['zhipu'] = {
|
||
'name': 'GLM-4-Flash',
|
||
'model_id': 'glm-4-flash',
|
||
'provider': 'zhipu',
|
||
'available': True
|
||
}
|
||
logger.info("智谱AI初始化成功 (使用模型: glm-4-flash)")
|
||
except Exception as e:
|
||
logger.error(f"智谱AI初始化失败: {e}")
|
||
|
||
# 初始化DeepSeek
|
||
if OPENAI_AVAILABLE and settings.deepseek_api_key:
|
||
try:
|
||
self.clients['deepseek'] = OpenAI(
|
||
api_key=settings.deepseek_api_key,
|
||
base_url="https://api.deepseek.com"
|
||
)
|
||
self.model_info['deepseek'] = {
|
||
'name': 'DeepSeek',
|
||
'model_id': 'deepseek-chat',
|
||
'provider': 'deepseek',
|
||
'available': True
|
||
}
|
||
logger.info("DeepSeek初始化成功")
|
||
except Exception as e:
|
||
logger.error(f"DeepSeek初始化失败: {e}")
|
||
|
||
# 设置默认模型（优先使用配置文件中的设置）
|
||
preferred_model = getattr(settings, 'smart_agent_model', None)
|
||
if preferred_model and preferred_model in self.clients:
|
||
self.current_model = preferred_model
|
||
logger.info(f"使用配置的模型: {preferred_model}")
|
||
elif 'deepseek' in self.clients:
|
||
self.current_model = 'deepseek'
|
||
elif 'zhipu' in self.clients:
|
||
self.current_model = 'zhipu'
|
||
|
||
if self.current_model:
|
||
logger.info(f"当前使用模型: {self.model_info[self.current_model]['name']}")
|
||
else:
|
||
logger.warning("没有可用的LLM模型")
|
||
|
||
def get_available_models(self) -> List[Dict[str, Any]]:
|
||
"""获取所有可用的模型列表"""
|
||
return [info for info in self.model_info.values() if info['available']]
|
||
|
||
def get_current_model_info(self) -> Optional[Dict[str, Any]]:
|
||
"""获取当前使用的模型信息"""
|
||
if self.current_model:
|
||
return self.model_info[self.current_model]
|
||
return None
|
||
|
||
def switch_model(self, provider: str) -> bool:
|
||
"""
|
||
切换模型
|
||
|
||
Args:
|
||
provider: 模型提供商 ('zhipu' 或 'deepseek')
|
||
|
||
Returns:
|
||
是否切换成功
|
||
"""
|
||
if provider in self.clients:
|
||
self.current_model = provider
|
||
logger.info(f"切换到模型: {self.model_info[provider]['name']}")
|
||
return True
|
||
else:
|
||
logger.error(f"模型不可用: {provider}")
|
||
return False
|
||
|
||
def _is_balance_error(self, error: Exception, provider: str) -> bool:
|
||
"""
|
||
检查错误是否是余额不足错误
|
||
|
||
Args:
|
||
error: 异常对象
|
||
provider: LLM提供商
|
||
|
||
Returns:
|
||
是否是余额不足错误
|
||
"""
|
||
error_str = str(error).lower()
|
||
error_type = type(error).__name__
|
||
|
||
# DeepSeek 余额错误
|
||
if provider == 'deepseek':
|
||
# APIStatusError: Error code: 402 - {'error': {'message': 'Insufficient Balance'
|
||
if '402' in error_str and 'insufficient balance' in error_str:
|
||
return True
|
||
if 'balance' in error_str and 'insufficient' in error_str:
|
||
return True
|
||
|
||
# 智谱AI 余额错误
|
||
elif provider == 'zhipu':
|
||
# 常见错误信息
|
||
if '余额' in error_str or 'balance' in error_str:
|
||
if 'insufficient' in error_str or '不足' in error_str:
|
||
return True
|
||
if error_type == 'APIError' and '130' in error_str: # 智谱错误码130表示余额不足
|
||
return True
|
||
|
||
return False
|
||
|
||
async def _notify_balance_error(self, provider: str, error: Exception):
|
||
"""
|
||
发送余额不足的Telegram通知
|
||
|
||
Args:
|
||
provider: LLM提供商
|
||
error: 异常对象
|
||
"""
|
||
# 检查冷却时间
|
||
now = datetime.now()
|
||
last_notified = self._balance_error_notified.get(provider)
|
||
if last_notified:
|
||
time_since_last = (now - last_notified).total_seconds()
|
||
if time_since_last < self.BALANCE_ERROR_COOLDOWN:
|
||
logger.info(f"{provider} 余额错误通知冷却中，剩余 {int(self.BALANCE_ERROR_COOLDOWN - time_since_last)} 秒")
|
||
return
|
||
|
||
# 发送通知
|
||
try:
|
||
from app.services.telegram_service import get_telegram_service
|
||
telegram = get_telegram_service()
|
||
|
||
provider_name = {
|
||
'zhipu': '智谱AI (GLM-4)',
|
||
'deepseek': 'DeepSeek'
|
||
}.get(provider, provider)
|
||
|
||
message = f"""🚨 <b>LLM API 余额不足警告</b>
|
||
|
||
━━━━━━━━━━━━━━━━━━━━
|
||
|
||
📊 <b>服务商:</b> {provider_name}
|
||
⚠️ <b>错误类型:</b> 余额不足 (Insufficient Balance)
|
||
🔍 <b>错误信息:</b> {str(error)[:200]}
|
||
|
||
━━━━━━━━━━━━━━━━━━━━
|
||
|
||
<i>请及时充值，否则智能体将无法正常工作</i>"""
|
||
|
||
await telegram.send_message(message, parse_mode="HTML")
|
||
logger.warning(f"已发送 {provider} 余额不足Telegram通知")
|
||
|
||
# 记录通知时间
|
||
self._balance_error_notified[provider] = now
|
||
|
||
except Exception as e:
|
||
logger.error(f"发送余额不足通知失败: {e}")
|
||
|
||
def chat(
|
||
self,
|
||
messages: List[Dict[str, str]],
|
||
temperature: float = 0.7,
|
||
max_tokens: int = 2000,
|
||
model_override: Optional[str] = None
|
||
) -> Optional[str]:
|
||
"""
|
||
调用LLM进行对话
|
||
|
||
Args:
|
||
messages: 消息列表
|
||
temperature: 温度参数
|
||
max_tokens: 最大token数
|
||
model_override: 临时覆盖使用的模型
|
||
|
||
Returns:
|
||
LLM响应文本
|
||
"""
|
||
provider = model_override or self.current_model
|
||
|
||
if not provider or provider not in self.clients:
|
||
logger.error("没有可用的LLM客户端")
|
||
return None
|
||
|
||
try:
|
||
client = self.clients[provider]
|
||
model_id = self.model_info[provider]['model_id']
|
||
|
||
logger.info(f"调用LLM: provider={provider}, model={model_id}, messages={len(messages)}条")
|
||
|
||
if provider == 'zhipu':
|
||
# 智谱AI调用
|
||
# Zhipu对参数更严格，temperature范围是0.0-1.0
|
||
safe_temperature = max(0.0, min(1.0, temperature))
|
||
logger.debug(f"智谱AI请求参数: model={model_id}, temperature={safe_temperature}, max_tokens={max_tokens}")
|
||
logger.debug(f"消息内容: {messages[-1]['content'][:200] if messages else 'empty'}...")
|
||
|
||
response = client.chat.completions.create(
|
||
model=model_id,
|
||
messages=messages,
|
||
temperature=safe_temperature,
|
||
max_tokens=max_tokens
|
||
)
|
||
logger.debug(f"智谱AI原始响应: {response}")
|
||
elif provider == 'deepseek':
|
||
# DeepSeek调用（OpenAI兼容）
|
||
# DeepSeek对参数更严格，确保temperature在有效范围内
|
||
safe_temperature = max(0.0, min(2.0, temperature))
|
||
response = client.chat.completions.create(
|
||
model=model_id,
|
||
messages=messages,
|
||
temperature=safe_temperature,
|
||
max_tokens=max_tokens
|
||
)
|
||
else:
|
||
logger.error(f"未知的模型提供商: {provider}")
|
||
return None
|
||
|
||
# 详细日志记录响应结构
|
||
logger.debug(f"响应对象类型: {type(response)}")
|
||
|
||
if hasattr(response, 'choices') and response.choices:
|
||
choice = response.choices[0]
|
||
logger.debug(f"Choice类型: {type(choice)}, 索引: {getattr(choice, 'index', 'N/A')}")
|
||
|
||
message = choice.message
|
||
logger.debug(f"Message类型: {type(message)}")
|
||
|
||
# 智谱AI可能使用不同的字段
|
||
content = getattr(message, 'content', None)
|
||
|
||
if content and content.strip():
|
||
logger.info(f"LLM响应成功，长度: {len(content)}")
|
||
return content
|
||
else:
|
||
logger.warning(f"LLM响应content为空或空白: content={repr(content)}")
|
||
# 尝试从其他字段获取内容（智谱AI可能使用不同字段）
|
||
for attr in ['reasoning_content', 'text', 'result']:
|
||
if hasattr(message, attr):
|
||
alt_content = getattr(message, attr)
|
||
if alt_content and alt_content.strip():
|
||
logger.info(f"从 {attr} 获取内容，长度: {len(alt_content)}")
|
||
return alt_content
|
||
|
||
# 打印完整的message对象用于调试
|
||
logger.debug(f"完整Message对象: {message}")
|
||
if hasattr(message, '__dict__'):
|
||
logger.debug(f"Message属性: {message.__dict__}")
|
||
|
||
return None
|
||
else:
|
||
logger.warning(f"LLM响应中没有choices。响应: {response}")
|
||
# 检查是否有其他可能的响应格式
|
||
if hasattr(response, 'content'):
|
||
content = response.content
|
||
if content and content.strip():
|
||
logger.info(f"从response.content获取内容，长度: {len(content)}")
|
||
return content
|
||
return None
|
||
|
||
except Exception as e:
|
||
logger.error(f"LLM调用失败: {type(e).__name__}: {e}")
|
||
import traceback
|
||
logger.error(f"详细错误: {traceback.format_exc()}")
|
||
|
||
# 检查是否是余额错误，发送Telegram通知
|
||
if self._is_balance_error(e, provider):
|
||
import asyncio
|
||
try:
|
||
# 在新的事件循环中运行（避免嵌套事件循环问题）
|
||
loop = asyncio.get_event_loop()
|
||
if loop.is_running():
|
||
# 如果在异步上下文中，创建任务
|
||
asyncio.create_task(self._notify_balance_error(provider, e))
|
||
else:
|
||
# 如果没有运行的循环，直接运行
|
||
asyncio.run(self._notify_balance_error(provider, e))
|
||
except Exception as notify_error:
|
||
logger.error(f"发送余额通知异常: {notify_error}")
|
||
|
||
return None
|
||
|
||
def chat_stream(
|
||
self,
|
||
messages: List[Dict[str, str]],
|
||
temperature: float = 0.7,
|
||
max_tokens: int = 2000,
|
||
model_override: Optional[str] = None
|
||
):
|
||
"""
|
||
流式调用LLM进行对话
|
||
|
||
Args:
|
||
messages: 消息列表
|
||
temperature: 温度参数
|
||
max_tokens: 最大token数
|
||
model_override: 临时覆盖使用的模型
|
||
|
||
Yields:
|
||
LLM响应的文本片段
|
||
"""
|
||
provider = model_override or self.current_model
|
||
|
||
if not provider or provider not in self.clients:
|
||
logger.error("没有可用的LLM客户端")
|
||
return
|
||
|
||
try:
|
||
client = self.clients[provider]
|
||
model_id = self.model_info[provider]['model_id']
|
||
|
||
logger.info(f"流式调用LLM: provider={provider}, model={model_id}, messages={len(messages)}条")
|
||
|
||
if provider == 'zhipu':
|
||
# 智谱AI流式调用
|
||
# Zhipu对参数更严格，temperature范围是0.0-1.0
|
||
safe_temperature = max(0.0, min(1.0, temperature))
|
||
response = client.chat.completions.create(
|
||
model=model_id,
|
||
messages=messages,
|
||
temperature=safe_temperature,
|
||
max_tokens=max_tokens,
|
||
stream=True
|
||
)
|
||
for chunk in response:
|
||
if chunk.choices and chunk.choices[0].delta.content:
|
||
yield chunk.choices[0].delta.content
|
||
|
||
elif provider == 'deepseek':
|
||
# DeepSeek流式调用（OpenAI兼容）
|
||
# DeepSeek对参数更严格，确保temperature在有效范围内
|
||
safe_temperature = max(0.0, min(2.0, temperature))
|
||
response = client.chat.completions.create(
|
||
model=model_id,
|
||
messages=messages,
|
||
temperature=safe_temperature,
|
||
max_tokens=max_tokens,
|
||
stream=True
|
||
)
|
||
for chunk in response:
|
||
if chunk.choices and chunk.choices[0].delta.content:
|
||
yield chunk.choices[0].delta.content
|
||
|
||
else:
|
||
logger.error(f"未知的模型提供商: {provider}")
|
||
return
|
||
|
||
logger.info("LLM流式响应完成")
|
||
|
||
except Exception as e:
|
||
logger.error(f"LLM流式调用失败: {type(e).__name__}: {e}")
|
||
import traceback
|
||
logger.error(f"详细错误: {traceback.format_exc()}")
|
||
|
||
# 检查是否是余额错误，发送Telegram通知
|
||
if self._is_balance_error(e, provider):
|
||
import asyncio
|
||
try:
|
||
# 在新的事件循环中运行（避免嵌套事件循环问题）
|
||
loop = asyncio.get_event_loop()
|
||
if loop.is_running():
|
||
# 如果在异步上下文中，创建任务
|
||
asyncio.create_task(self._notify_balance_error(provider, e))
|
||
else:
|
||
# 如果没有运行的循环，直接运行
|
||
asyncio.run(self._notify_balance_error(provider, e))
|
||
except Exception as notify_error:
|
||
logger.error(f"发送余额通知异常: {notify_error}")
|
||
|
||
return
|
||
|
||
def analyze_intent(self, user_message: str) -> Dict[str, Any]:
|
||
"""使用LLM分析用户意图"""
|
||
if not self.current_model:
|
||
return {"type": "unknown", "confidence": 0}
|
||
|
||
prompt = f"""你是一个股票分析助手的意图识别模块。请分析用户的查询意图。
|
||
|
||
用户消息：{user_message}
|
||
|
||
请识别以下意图类型之一：
|
||
1. market_data - 查询实时行情、价格
|
||
2. technical_analysis - 技术分析、技术指标
|
||
3. fundamental - 基本面信息、公司信息
|
||
4. visualization - K线图、图表
|
||
5. unknown - 无法识别
|
||
|
||
请以JSON格式返回：
|
||
{{
|
||
"type": "意图类型",
|
||
"confidence": 0.0-1.0,
|
||
"stock_name": "提取的股票名称（如果有）"
|
||
}}
|
||
"""
|
||
|
||
try:
|
||
response = self.chat([{"role": "user", "content": prompt}], temperature=0.3)
|
||
if response:
|
||
import json
|
||
result = json.loads(response)
|
||
return result
|
||
except Exception as e:
|
||
logger.error(f"意图分析失败: {e}")
|
||
|
||
return {"type": "unknown", "confidence": 0}
|
||
|
||
|
||
# 创建全局实例
|
||
multi_llm_service = MultiLLMService()
|