Eric Wang
e5d2a78886
- Config.llm_profiles: LLM_PROVIDERS declares names, each group defined by LLM_<NAME>_BASE_URL/_API_KEY/_MODEL (required) and optional _TIMEOUT_SECONDS/ _MAX_RETRIES/_TEMPERATURE/_EXTRA_BODY; JSON-dict form of LLM_PROVIDERS also supported. LLM_PROVIDER selects the active group at runtime; empty/fake = offline scripted model - removed fixed openai/kimi groups (AGENT_LLM_PROVIDER, KIMI_LLM_*) - validate(): incomplete groups and unknown LLM_PROVIDER fail fast with the configured-name list - OpenAILLMService: http client trust_env from LLM_HTTP_TRUST_ENV (default False - local gateways must bypass the Windows system proxy) - build_llm_service(provider): explicit name or LLM_PROVIDER default - .env/.env.example: deepseek + kimi groups side by side, switch by LLM_PROVIDER=kimi (thinking off via EXTRA_BODY reasoning_effort=none) - eval probe scripts updated to the new config interface Verified: 172 unit tests; live calls succeed on both providers (deepseek 28s, kimi 6.6s, both emitting valid <state> prefixes)
276 lines
11 KiB
Python
276 lines
11 KiB
Python
"""LLM 服务（作业 4）。
|
||
|
||
fake 与真实 provider 实现同一 ``LLMService`` Protocol，节点不感知差异：
|
||
- ``FakeLLMService``：脚本化固定输出，离线可重复，日常测试用它；
|
||
- ``OpenAILLMService``：真实 provider（TODO 关键代码）——
|
||
超时、取消、有限重试、模型配置。
|
||
|
||
``build_llm_service`` 按命名提供商列表（Config.LLM_PROVIDERS / LLM_PROVIDER）
|
||
选择实现，见 core.config。
|
||
"""
|
||
|
||
from collections.abc import Awaitable, Callable
|
||
from typing import Protocol
|
||
|
||
Responder = Callable[[str, str], str | Awaitable[str]]
|
||
|
||
|
||
class LLMError(Exception):
|
||
"""LLM 调用失败基类。"""
|
||
|
||
|
||
class LLMTimeoutError(LLMError):
|
||
"""调用超过超时时间。"""
|
||
|
||
|
||
class LLMAPIError(LLMError):
|
||
"""provider 返回错误或无效响应。"""
|
||
|
||
|
||
class LLMService(Protocol):
|
||
"""节点依赖的最小 LLM 接口。"""
|
||
|
||
async def complete(
|
||
self, system_prompt: str, user_message: str, *,
|
||
history: list[dict[str, str]] | None = None,
|
||
) -> str:
|
||
"""返回模型文本；history 保留 user/assistant 角色，与工作流对齐。"""
|
||
...
|
||
|
||
|
||
class FakeLLMService:
|
||
"""按 responder 返回固定输出的假 LLM。
|
||
|
||
responder 接收 (system_prompt, user_message)，返回 str（或
|
||
coroutine）。每次调用记录在 ``calls``，测试用它断言
|
||
“一次轮次只调用一次模型”等约束。
|
||
"""
|
||
|
||
def __init__(self, responder: Responder):
|
||
self._responder = responder
|
||
self.calls: list[dict] = []
|
||
# 兼容两参 responder；三参 responder 可读取对话历史。
|
||
import inspect
|
||
|
||
self._accepts_history = (
|
||
len(inspect.signature(responder).parameters) >= 3
|
||
)
|
||
|
||
async def complete(
|
||
self, system_prompt: str, user_message: str, *,
|
||
history: list[dict[str, str]] | None = None,
|
||
) -> str:
|
||
self.calls.append(
|
||
{"system_prompt": system_prompt, "user_message": user_message,
|
||
"history": [dict(message) for message in (history or [])]}
|
||
)
|
||
import asyncio
|
||
|
||
result = (
|
||
self._responder(system_prompt, user_message, history or [])
|
||
if self._accepts_history
|
||
else self._responder(system_prompt, user_message)
|
||
)
|
||
if asyncio.iscoroutine(result):
|
||
result = await result
|
||
return result
|
||
|
||
|
||
class OpenAILLMService:
|
||
"""真实 LLM provider（OpenAI 兼容接口，如 OpenAI / DeepSeek 等）。
|
||
|
||
参考实现：async 客户端 + asyncio.wait_for 超时（取消会向外传播，
|
||
客户端断连时图任务随之取消）；仅对可重试错误（429 / 5xx / 连接错）
|
||
按 LLM_MAX_RETRIES 指数退避重试；错误统一转换为
|
||
LLMTimeoutError / LLMAPIError，不向上层泄漏原始响应体。
|
||
"""
|
||
|
||
def __init__(
|
||
self,
|
||
model: str,
|
||
api_key: str,
|
||
base_url: str = "",
|
||
timeout_seconds: float = 30.0,
|
||
max_retries: int = 2,
|
||
temperature: float | None = None,
|
||
extra_body: dict | None = None,
|
||
):
|
||
self._model = model
|
||
self._api_key = api_key
|
||
self._base_url = base_url
|
||
self._timeout_seconds = timeout_seconds
|
||
self._max_retries = max_retries
|
||
self._temperature = temperature
|
||
# 附加请求体（如思考开关），原样合并进 chat.completions.create 参数。
|
||
self._extra_body = dict(extra_body or {})
|
||
self._client = None
|
||
|
||
def _get_client(self):
|
||
# 延迟构造：fake provider 使用时无需安装/初始化真实客户端。
|
||
# AsyncOpenAI 的创建不绑定事件循环，可在首次调用前安全缓存。
|
||
# trust_env=False：不走系统代理，本地网关直连（代理曾导致 503/500）。
|
||
if self._client is None:
|
||
import httpx
|
||
from openai import AsyncOpenAI
|
||
|
||
from ..core.config import Config
|
||
|
||
kwargs = {"api_key": self._api_key, "timeout": self._timeout_seconds}
|
||
if self._base_url:
|
||
kwargs["base_url"] = self._base_url
|
||
kwargs["http_client"] = httpx.AsyncClient(trust_env=Config.LLM_HTTP_TRUST_ENV)
|
||
self._client = AsyncOpenAI(**kwargs)
|
||
return self._client
|
||
|
||
async def complete(
|
||
self, system_prompt: str, user_message: str, *,
|
||
history: list[dict[str, str]] | None = None,
|
||
) -> str:
|
||
import asyncio
|
||
|
||
from openai import APIConnectionError, APITimeoutError, RateLimitError
|
||
|
||
client = self._get_client()
|
||
messages = [
|
||
{"role": "system", "content": system_prompt},
|
||
*[dict(message) for message in (history or [])],
|
||
{"role": "user", "content": user_message},
|
||
]
|
||
kwargs: dict = {"model": self._model, "messages": messages}
|
||
if self._temperature is not None:
|
||
kwargs["temperature"] = self._temperature
|
||
kwargs.update(self._extra_body)
|
||
retries = 0
|
||
while True:
|
||
try:
|
||
response = await asyncio.wait_for(
|
||
client.chat.completions.create(**kwargs),
|
||
timeout=self._timeout_seconds,
|
||
)
|
||
except (APITimeoutError, asyncio.TimeoutError) as exc:
|
||
# 超时：可重试，重试耗尽后转换为 LLMTimeoutError
|
||
if retries < self._max_retries:
|
||
retries += 1
|
||
await asyncio.sleep(min(2 ** retries, 10))
|
||
continue
|
||
raise LLMTimeoutError(
|
||
f"LLM 调用超时（{self._timeout_seconds}s，"
|
||
f"重试 {self._max_retries} 次后仍失败）"
|
||
) from exc
|
||
except (RateLimitError, APIConnectionError) as exc:
|
||
# 限流 / 连接错误：可重试
|
||
if retries < self._max_retries:
|
||
retries += 1
|
||
await asyncio.sleep(min(2 ** retries, 10))
|
||
continue
|
||
raise LLMAPIError(f"LLM provider 不可用：{type(exc).__name__}") from exc
|
||
except Exception as exc:
|
||
# 其他错误（如 4xx 参数错）不可重试
|
||
raise LLMAPIError(f"LLM 调用失败：{type(exc).__name__}") from exc
|
||
|
||
content = response.choices[0].message.content if response.choices else None
|
||
if not content:
|
||
raise LLMAPIError("LLM 返回了空内容")
|
||
return content
|
||
|
||
|
||
def _default_fake_responder(system_prompt: str, user_message: str) -> str:
|
||
"""离线兜底假回复：回显输入，保持服务在无配置时可启动。"""
|
||
return f"<state>1002</state>已收到：{user_message}"
|
||
|
||
|
||
def workflow_fake_responder(system_prompt: str, user_message: str, history=None) -> str:
|
||
"""模拟工作流提示词行为的假 LLM（离线测试/冒烟用）。
|
||
|
||
从注入的运行时上下文（system 末尾的 context_prompt）读取当前
|
||
阶段与 clientMode；按工作流提示词语义响应客户端事件：
|
||
会话标记 1001→1002、【拍摄完成】沿链推进、失败一次→0005、
|
||
第二次无回复→0004；普通输入保持当前阶段。正文字面回显输入。
|
||
"""
|
||
import re
|
||
|
||
from .events import ClientEvent, parse_client_event
|
||
|
||
stage_match = re.search(r'"stage_code":\s*"(\d{4})"', system_prompt)
|
||
current = stage_match.group(1) if stage_match else "1001"
|
||
mode = "browser_addon" if '"client_mode": "browser_addon"' in system_prompt else "direct"
|
||
|
||
# 20261022 起 user 消息为 <turn_context>+<user_input> 包装，取原始输入判断事件
|
||
input_match = re.search(r"<user_input>\n(.*?)\n?</user_input>", user_message, re.DOTALL)
|
||
raw_input = input_match.group(1) if input_match else user_message
|
||
|
||
event = parse_client_event(raw_input)
|
||
if event in (ClientEvent.SESSION_START, ClientEvent.CONTINUE) and current == "1001":
|
||
return f"<state>1002</state>已收到：{raw_input}"
|
||
if event is ClientEvent.PHOTO_DONE:
|
||
chain = {
|
||
"2000": "2001", "2001": "2002", "2002": "2003", "2003": "2005",
|
||
"2010": "2011", "2011": "2012", "2012": "2013",
|
||
"2013": "2014", "2014": "2015",
|
||
}
|
||
target = chain.get(current)
|
||
if target is None:
|
||
target = "0000" if mode == "browser_addon" else ("3001" if current == "2005" else "3002")
|
||
return f"<state>{target}</state>已收到：{raw_input}"
|
||
if event is ClientEvent.PHOTO_FAILED_3X:
|
||
return "<state>0005</state>多次拍摄识别仍未成功，正在为您转接人工处理。"
|
||
if event is ClientEvent.NO_REPLY:
|
||
trailing = 0
|
||
for message in reversed(history or []):
|
||
if message.get("role") != "user":
|
||
continue
|
||
if parse_client_event(str(message.get("content", ""))) is ClientEvent.NO_REPLY:
|
||
trailing += 1
|
||
else:
|
||
break
|
||
if trailing >= 1:
|
||
return "<state>0004</state>长时间未收到回应，为您转接人工。"
|
||
return f"<state>{current}</state>请问您还在吗？"
|
||
return f"<state>{current}</state>已收到：{raw_input}"
|
||
|
||
|
||
def workflow_fake_llm() -> "FakeLLMService":
|
||
"""离线假 LLM：行为对齐工作流提示词，供冒烟与无模型测试。"""
|
||
return FakeLLMService(workflow_fake_responder)
|
||
|
||
|
||
def build_llm_service(provider: str | None = None) -> LLMService:
|
||
"""按命名提供商列表构造 LLM 服务；provider 缺省读 Config.LLM_PROVIDER。
|
||
|
||
配置见 Config.llm_profiles（LLM_PROVIDERS 名单 + LLM_<名>_* 变量）。
|
||
provider 为空或 fake = 离线脚本模型。
|
||
"""
|
||
from ..core.config import Config
|
||
|
||
provider = (provider if provider is not None else Config.LLM_PROVIDER)
|
||
provider = (provider or "").strip().lower()
|
||
if not provider or provider == "fake":
|
||
return FakeLLMService(_default_fake_responder)
|
||
profiles = Config.llm_profiles()
|
||
if provider not in profiles:
|
||
raise ValueError(
|
||
f"Unknown LLM provider {provider!r}; configured: {sorted(profiles)}"
|
||
)
|
||
profile = profiles[provider]
|
||
extra_body: dict = {}
|
||
if profile["extra_body"]:
|
||
import json as _json
|
||
|
||
try:
|
||
parsed = _json.loads(profile["extra_body"])
|
||
if isinstance(parsed, dict):
|
||
extra_body = parsed
|
||
except ValueError as exc:
|
||
raise ValueError(
|
||
f"extra_body of provider {provider!r} is not valid JSON: {exc}"
|
||
) from exc
|
||
return OpenAILLMService(
|
||
model=profile["model"],
|
||
api_key=profile["api_key"],
|
||
base_url=profile["base_url"],
|
||
timeout_seconds=profile["timeout_seconds"],
|
||
max_retries=profile["max_retries"],
|
||
temperature=profile["temperature"],
|
||
extra_body=extra_body,
|
||
)
|