| 123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175 |
- import aiohttp
- import asyncio
- import json
- import os
- import re
- from typing import Optional, Dict, Any
- from sqlalchemy.ext.asyncio import AsyncSession
- from sqlalchemy import select
- from app.core.config import settings
- from app.core.biz_exception import NotFoundError, BizLogicError
- from app.core.logger import logger
- from app.schemas.llm import ParseUserInputsPayload, ParseUserInputsOut
- from app.models.schema import VasSchema
- class LlmService:
- """
- 通用 LLM 服务(OpenAI-compatible 协议)
- 支持 Qwen (DashScope)、OpenAI、DeepSeek、Moonshot、Ollama 等任意标准兼容模型
- """
- @classmethod
- def get_api_key(cls) -> str:
- """获取 API Key(优先读取 settings.llm_api_key,其次读取环境变量 LLM_API_KEY)"""
- return (
- (settings.llm_api_key or "").strip()
- or os.getenv("LLM_API_KEY", "").strip()
- )
- @classmethod
- def get_completions_url(cls) -> str:
- """获取并规范化 /chat/completions 完整请求 URL"""
- base = (
- (settings.llm_base_url or "").strip()
- or os.getenv("LLM_BASE_URL", "https://dashscope.aliyuncs.com/compatible-mode/v1").strip()
- ).rstrip("/")
-
- if base.endswith("/chat/completions"):
- return base
- return f"{base}/chat/completions"
- @classmethod
- def get_default_model(cls) -> str:
- """获取默认模型名称"""
- return (
- (settings.llm_model or "").strip()
- or os.getenv("LLM_MODEL", "qwen-plus").strip()
- )
- @classmethod
- def clean_json_response(cls, content_str: str) -> Dict[str, Any]:
- """
- 从 LLM 返回的内容中提取并解析合法 JSON,自动剥离 Markdown 代码块
- """
- if not content_str:
- return {}
- text = content_str.strip()
- # 匹配 ```json ... ``` 或 ``` ... ```
- code_block_match = re.search(r"```(?:json)?\s*([\s\S]*?)\s*```", text, re.IGNORECASE)
- if code_block_match:
- text = code_block_match.group(1).strip()
- # 尝试直接解析
- try:
- return json.loads(text)
- except json.JSONDecodeError:
- # 尝试提取最外层的 { ... }
- brace_match = re.search(r"(\{[\s\S]*\})", text)
- if brace_match:
- return json.loads(brace_match.group(1))
- raise
- @classmethod
- async def handle_parse(cls, db: AsyncSession, payload: ParseUserInputsPayload) -> ParseUserInputsOut:
- """
- 处理数据解析接口请求
- """
- stmt = select(VasSchema).where(VasSchema.id == payload.schema_id)
- obj = (await db.execute(stmt)).scalar_one_or_none()
- if not obj:
- raise NotFoundError("Schema not exist")
- parsed_obj = await cls.parse_data_async(
- user_text=payload.input_raw_str,
- json_schema=obj.schema_json,
- model=payload.model
- )
- return ParseUserInputsOut(parsed_obj=parsed_obj)
- @classmethod
- async def parse_data_async(
- cls,
- user_text: str,
- json_schema: dict,
- model: Optional[str] = None,
- max_retries: int = 1
- ) -> Dict[str, Any]:
- """
- [异步版本] 调用配置好的 OpenAI 兼容接口解析数据
- """
- api_key = cls.get_api_key()
- if not api_key:
- logger.warning("[LlmService] No LLM API Key configured.")
- raise BizLogicError("LLM API Key not configured")
- url = cls.get_completions_url()
- selected_model = (model or "").strip() or cls.get_default_model()
- temperature = settings.llm_temperature
- timeout_seconds = settings.llm_timeout_seconds or 30
- headers = {
- "Authorization": f"Bearer {api_key}",
- "Content-Type": "application/json",
- }
- # 构造 Prompt
- system_instruction = "You are a specialized data extraction API. Output valid JSON only."
- user_prompt = f"""
- Extract data from the text strictly based on the provided JSON Schema.
- [JSON Schema]
- {json.dumps(json_schema, ensure_ascii=False)}
- [User Text]
- {user_text}
- """
- payload = {
- "model": selected_model,
- "messages": [
- {"role": "system", "content": system_instruction},
- {"role": "user", "content": user_prompt}
- ],
- "temperature": temperature,
- "response_format": {"type": "json_object"}
- }
- timeout = aiohttp.ClientTimeout(total=float(timeout_seconds))
- for attempt in range(max_retries + 1):
- try:
- async with aiohttp.ClientSession(timeout=timeout) as session:
- async with session.post(url, headers=headers, json=payload) as response:
- if response.status != 200:
- error_text = await response.text()
- logger.error(f"[LlmService] HTTP {response.status} from {url}: {error_text}")
- if attempt == max_retries:
- raise BizLogicError(f"LLM request failed: HTTP {response.status}")
- await asyncio.sleep(0.5)
- continue
- result = await response.json()
- choices = result.get("choices", [])
- if not choices:
- raise BizLogicError("LLM response did not contain completion choices")
- content_str = choices[0].get("message", {}).get("content", "")
- try:
- return cls.clean_json_response(content_str)
- except Exception as e:
- logger.error(f"[LlmService] Failed to parse JSON from LLM response: {content_str}, error: {e}")
- raise BizLogicError("Failed to parse valid JSON from LLM response")
- except BizLogicError:
- raise
- except Exception as e:
- logger.error(f"[LlmService] Exception on attempt {attempt + 1}: {e}")
- if attempt == max_retries:
- raise BizLogicError(f"LLM invocation error: {str(e)}")
- await asyncio.sleep(0.5)
- raise BizLogicError("LLM request failed")
|