281 lines
12 KiB
Python
281 lines
12 KiB
Python
|
import asyncio
|
||
|
|
import json
|
||
|
|
import time
|
||
|
|
from collections.abc import AsyncIterator
|
||
|
|
from dataclasses import dataclass
|
||
|
|
|
||
|
|
import httpx
|
||
|
|
|
||
|
|
from app.core.config import settings
|
||
|
|
from app.core.exceptions import AppError
|
||
|
|
|
||
|
|
|
||
|
|
@dataclass
|
||
|
|
class LLMResponse:
|
||
|
|
"""LLM 响应：封装非流式模型输出和耗时指标。"""
|
||
|
|
|
||
|
|
content: str
|
||
|
|
model: str
|
||
|
|
latency_ms: int
|
||
|
|
token_usage: dict | None = None
|
||
|
|
|
||
|
|
|
||
|
|
@dataclass
|
||
|
|
class LLMStreamChunk:
|
||
|
|
"""LLM 流式片段：封装 SSE 增量内容和完成状态。"""
|
||
|
|
|
||
|
|
delta: str
|
||
|
|
done: bool = False
|
||
|
|
first_token_ms: int | None = None
|
||
|
|
total_latency_ms: int | None = None
|
||
|
|
model: str | None = None
|
||
|
|
fallback_used: bool = False
|
||
|
|
|
||
|
|
|
||
|
|
class OpenAICompatibleLLMClient:
|
||
|
|
"""LLM Adapter：统一封装 OpenAI-compatible 模型的可替换调用。"""
|
||
|
|
|
||
|
|
def __init__(self) -> None:
|
||
|
|
self.base_url = settings.llm_base_url.rstrip("/")
|
||
|
|
self.api_key = settings.llm_api_key
|
||
|
|
self.timeout = settings.llm_timeout_seconds
|
||
|
|
self.chat_completions_url = self._build_chat_completions_url()
|
||
|
|
|
||
|
|
@property
|
||
|
|
def is_mock_mode(self) -> bool:
|
||
|
|
"""模型模式：没有 API Key 或开启 mock 时使用本地模拟响应。"""
|
||
|
|
return settings.llm_mock_enabled or not self.api_key
|
||
|
|
|
||
|
|
async def chat(
|
||
|
|
self,
|
||
|
|
messages: list[dict],
|
||
|
|
model: str,
|
||
|
|
*,
|
||
|
|
thinking_enabled: bool | None = None,
|
||
|
|
reasoning_effort: str | None = None,
|
||
|
|
response_format: dict | None = None,
|
||
|
|
max_tokens: int | None = None,
|
||
|
|
) -> LLMResponse:
|
||
|
|
"""非流式调用：向 OpenAI-compatible 接口发送 messages 并返回完整文本。"""
|
||
|
|
start = time.perf_counter()
|
||
|
|
if self.is_mock_mode:
|
||
|
|
content = self._mock_response(messages)
|
||
|
|
return LLMResponse(content=content, model=f"mock-{model}", latency_ms=int((time.perf_counter() - start) * 1000))
|
||
|
|
|
||
|
|
try:
|
||
|
|
async with httpx.AsyncClient(timeout=self._http_timeout()) as client:
|
||
|
|
resp = await client.post(
|
||
|
|
self.chat_completions_url,
|
||
|
|
headers={"Authorization": f"Bearer {self.api_key}"},
|
||
|
|
json=self._build_payload(
|
||
|
|
model=model,
|
||
|
|
messages=messages,
|
||
|
|
stream=False,
|
||
|
|
thinking_enabled=thinking_enabled,
|
||
|
|
reasoning_effort=reasoning_effort,
|
||
|
|
response_format=response_format,
|
||
|
|
max_tokens=max_tokens,
|
||
|
|
),
|
||
|
|
)
|
||
|
|
resp.raise_for_status()
|
||
|
|
data = resp.json()
|
||
|
|
content = (data["choices"][0]["message"].get("content") or "").strip()
|
||
|
|
if not content:
|
||
|
|
raise KeyError("empty llm content")
|
||
|
|
return LLMResponse(
|
||
|
|
content=content,
|
||
|
|
model=model,
|
||
|
|
latency_ms=int((time.perf_counter() - start) * 1000),
|
||
|
|
token_usage=data.get("usage"),
|
||
|
|
)
|
||
|
|
except (httpx.TimeoutException, httpx.HTTPError, KeyError, IndexError, json.JSONDecodeError) as exc:
|
||
|
|
if settings.llm_fallback_to_mock:
|
||
|
|
content = self._mock_response(messages)
|
||
|
|
return LLMResponse(
|
||
|
|
content=content,
|
||
|
|
model=f"mock-fallback-{model}",
|
||
|
|
latency_ms=int((time.perf_counter() - start) * 1000),
|
||
|
|
token_usage={"fallback_reason": exc.__class__.__name__},
|
||
|
|
)
|
||
|
|
raise AppError("LLM_CALL_FAILED", "llm service call failed", 502) from exc
|
||
|
|
|
||
|
|
async def stream_chat(
|
||
|
|
self,
|
||
|
|
messages: list[dict],
|
||
|
|
model: str,
|
||
|
|
*,
|
||
|
|
thinking_enabled: bool | None = None,
|
||
|
|
reasoning_effort: str | None = None,
|
||
|
|
max_tokens: int | None = None,
|
||
|
|
) -> AsyncIterator[LLMStreamChunk]:
|
||
|
|
"""流式调用：以统一 chunk 结构输出 OpenAI-compatible SSE 增量。"""
|
||
|
|
start = time.perf_counter()
|
||
|
|
first_token_ms: int | None = None
|
||
|
|
if self.is_mock_mode:
|
||
|
|
async for chunk in self._mock_stream(messages, model, start, model_label=f"mock-{model}"):
|
||
|
|
yield chunk
|
||
|
|
return
|
||
|
|
|
||
|
|
try:
|
||
|
|
async with httpx.AsyncClient(timeout=self._http_timeout()) as client:
|
||
|
|
async with client.stream(
|
||
|
|
"POST",
|
||
|
|
self.chat_completions_url,
|
||
|
|
headers={"Authorization": f"Bearer {self.api_key}"},
|
||
|
|
json=self._build_payload(
|
||
|
|
model=model,
|
||
|
|
messages=messages,
|
||
|
|
stream=True,
|
||
|
|
thinking_enabled=thinking_enabled,
|
||
|
|
reasoning_effort=reasoning_effort,
|
||
|
|
max_tokens=max_tokens,
|
||
|
|
),
|
||
|
|
) as resp:
|
||
|
|
resp.raise_for_status()
|
||
|
|
async for line in resp.aiter_lines():
|
||
|
|
if not line.startswith("data:"):
|
||
|
|
continue
|
||
|
|
payload = line.removeprefix("data:").strip()
|
||
|
|
if payload == "[DONE]":
|
||
|
|
break
|
||
|
|
data = json.loads(payload)
|
||
|
|
delta_obj = data["choices"][0].get("delta", {})
|
||
|
|
content_delta = delta_obj.get("content") or ""
|
||
|
|
reasoning_delta = delta_obj.get("reasoning_content") or ""
|
||
|
|
if (content_delta or reasoning_delta) and first_token_ms is None:
|
||
|
|
first_token_ms = int((time.perf_counter() - start) * 1000)
|
||
|
|
if content_delta:
|
||
|
|
yield LLMStreamChunk(delta=content_delta, first_token_ms=first_token_ms)
|
||
|
|
except (httpx.TimeoutException, httpx.HTTPError, KeyError, IndexError, json.JSONDecodeError) as exc:
|
||
|
|
if settings.llm_fallback_to_mock:
|
||
|
|
async for chunk in self._mock_stream(
|
||
|
|
messages,
|
||
|
|
model,
|
||
|
|
start,
|
||
|
|
model_label=f"mock-fallback-{model}",
|
||
|
|
fallback_used=True,
|
||
|
|
):
|
||
|
|
yield chunk
|
||
|
|
return
|
||
|
|
raise AppError("LLM_STREAM_FAILED", "llm stream call failed", 502) from exc
|
||
|
|
|
||
|
|
yield LLMStreamChunk(
|
||
|
|
delta="",
|
||
|
|
done=True,
|
||
|
|
first_token_ms=first_token_ms,
|
||
|
|
total_latency_ms=int((time.perf_counter() - start) * 1000),
|
||
|
|
model=model,
|
||
|
|
)
|
||
|
|
|
||
|
|
async def _mock_stream(
|
||
|
|
self,
|
||
|
|
messages: list[dict],
|
||
|
|
model: str,
|
||
|
|
start: float,
|
||
|
|
model_label: str,
|
||
|
|
fallback_used: bool = False,
|
||
|
|
) -> AsyncIterator[LLMStreamChunk]:
|
||
|
|
"""Mock 流式输出：在模型不可用时保持 Demo 流程可验证。"""
|
||
|
|
first_token_ms: int | None = None
|
||
|
|
content = self._mock_response(messages)
|
||
|
|
for piece in self._split_mock_content(content):
|
||
|
|
await asyncio.sleep(0.02)
|
||
|
|
if first_token_ms is None:
|
||
|
|
first_token_ms = int((time.perf_counter() - start) * 1000)
|
||
|
|
yield LLMStreamChunk(delta=piece, first_token_ms=first_token_ms)
|
||
|
|
yield LLMStreamChunk(
|
||
|
|
delta="",
|
||
|
|
done=True,
|
||
|
|
first_token_ms=first_token_ms,
|
||
|
|
total_latency_ms=int((time.perf_counter() - start) * 1000),
|
||
|
|
model=model_label,
|
||
|
|
fallback_used=fallback_used,
|
||
|
|
)
|
||
|
|
|
||
|
|
def _mock_response(self, messages: list[dict]) -> str:
|
||
|
|
"""Mock 输出：在没有 DeepSeek Key 时保证 Demo 闭环可运行。"""
|
||
|
|
latest = next((m.get("content", "") for m in reversed(messages) if m.get("role") == "user"), "")
|
||
|
|
prompt_head = " ".join(m.get("content", "").lower() for m in messages[:2])
|
||
|
|
if "score_type" in prompt_head and "dimension_scores" in prompt_head:
|
||
|
|
return json.dumps(
|
||
|
|
{
|
||
|
|
"score_type": "percentage",
|
||
|
|
"total_score": 82,
|
||
|
|
"dimension_scores": [
|
||
|
|
{"dimension": "信息获取", "score": 20, "max_score": 25, "comment": "覆盖了发热、咳嗽和喘息，儿科特异性病史仍需加强。"},
|
||
|
|
{"dimension": "分析推理", "score": 21, "max_score": 25, "comment": "能够识别肺炎方向，鉴别诊断完整性中等。"},
|
||
|
|
{"dimension": "处置决策", "score": 17, "max_score": 20, "comment": "治疗原则基本合理，风险预案需要更具体。"},
|
||
|
|
{"dimension": "沟通人文", "score": 12, "max_score": 15, "comment": "有告知意识，家属安抚和健康教育可更系统。"},
|
||
|
|
{"dimension": "临床整合", "score": 12, "max_score": 15, "comment": "诊疗流程完整，时间分配和整体组织较清晰。"},
|
||
|
|
],
|
||
|
|
"errors": [{"title": "儿科特异性病史不足", "description": "疫苗接种、过敏史、既往喘息史追问不足。"}],
|
||
|
|
"improvement_plan": ["补充儿科问诊框架：出生史、接种史、过敏史、既往喘息史。"],
|
||
|
|
"evidence_summary": ["用户完成了核心症状追问、检查申请、诊断和治疗提交。"],
|
||
|
|
"guideline_refs": [],
|
||
|
|
"overall_comment": "本次训练完成主要诊疗流程，诊断方向正确，治疗方案具备基本可执行性。",
|
||
|
|
},
|
||
|
|
ensure_ascii=False,
|
||
|
|
)
|
||
|
|
if "体温" in latest or "发热" in latest:
|
||
|
|
return "最高烧到39度多，已经反复四天了，退烧后会好一点，但很快又起来。"
|
||
|
|
if "喘" in latest or "呼吸" in latest:
|
||
|
|
return "昨天开始喘得明显，活动后更明显，晚上咳嗽也更重。"
|
||
|
|
if "精神" in latest or "吃" in latest:
|
||
|
|
return "精神比平时差一些，吃饭少了，但还能喝水，小便比平时略少。"
|
||
|
|
if "既往" in latest or "过敏" in latest:
|
||
|
|
return "以前没有明确哮喘诊断，也没有药物过敏史，小时候感冒时偶尔会咳得久。"
|
||
|
|
return "家长：孩子主要是发热、咳嗽，昨天开始喘，您可以继续问我具体情况。"
|
||
|
|
|
||
|
|
def _split_mock_content(self, content: str) -> list[str]:
|
||
|
|
"""Mock 分片：把本地模拟文本拆成流式输出片段。"""
|
||
|
|
return [content[i : i + 8] for i in range(0, len(content), 8)]
|
||
|
|
|
||
|
|
def _build_chat_completions_url(self) -> str:
|
||
|
|
"""接口地址：兼容 base_url 和完整 chat/completions URL 两种写法。"""
|
||
|
|
if self.base_url.endswith("/chat/completions"):
|
||
|
|
return self.base_url
|
||
|
|
return f"{self.base_url}/chat/completions"
|
||
|
|
|
||
|
|
def _http_timeout(self) -> httpx.Timeout:
|
||
|
|
"""超时策略：限制连接、写入和读取等待，避免前端长时间卡在生成中。"""
|
||
|
|
return httpx.Timeout(
|
||
|
|
timeout=self.timeout,
|
||
|
|
connect=min(8, self.timeout),
|
||
|
|
read=self.timeout,
|
||
|
|
write=min(15, self.timeout),
|
||
|
|
pool=min(8, self.timeout),
|
||
|
|
)
|
||
|
|
|
||
|
|
def _build_payload(
|
||
|
|
self,
|
||
|
|
*,
|
||
|
|
model: str,
|
||
|
|
messages: list[dict],
|
||
|
|
stream: bool,
|
||
|
|
thinking_enabled: bool | None = None,
|
||
|
|
reasoning_effort: str | None = None,
|
||
|
|
response_format: dict | None = None,
|
||
|
|
max_tokens: int | None = None,
|
||
|
|
) -> dict:
|
||
|
|
"""请求构造：兼容 DeepSeek V4 thinking、reasoning_effort 和 JSON 输出。"""
|
||
|
|
payload: dict = {"model": model, "messages": messages, "stream": stream}
|
||
|
|
supports_reasoning_options = self._supports_reasoning_options(model)
|
||
|
|
if thinking_enabled is not None and supports_reasoning_options:
|
||
|
|
payload["thinking"] = {"type": "enabled" if thinking_enabled else "disabled"}
|
||
|
|
if reasoning_effort and supports_reasoning_options and thinking_enabled is not False:
|
||
|
|
payload["reasoning_effort"] = reasoning_effort
|
||
|
|
if response_format:
|
||
|
|
payload["response_format"] = response_format
|
||
|
|
if max_tokens:
|
||
|
|
payload["max_tokens"] = max_tokens
|
||
|
|
return payload
|
||
|
|
|
||
|
|
def _supports_reasoning_options(self, model: str) -> bool:
|
||
|
|
"""厂商兼容：只向 DeepSeek 发送 thinking/reasoning_effort 等专有参数。"""
|
||
|
|
base = self.base_url.lower()
|
||
|
|
model_name = model.lower()
|
||
|
|
return "deepseek" in base or model_name.startswith("deepseek")
|
||
|
|
|
||
|
|
|
||
|
|
DeepSeekClient = OpenAICompatibleLLMClient
|