放弃探针方案,但协程独立重试

This commit is contained in:
zk
2026-07-09 19:21:35 +08:00
parent 64bf6d7af0
commit dfb0f177e2
+7 -59
View File
@@ -1,8 +1,7 @@
"""AI 调用工具封装
核心机制:健康门闸(Health Gate)
- ainvoke 抛异常 → 标记故障,全部协程阻塞,单探针退避重试直到成功
- ainvoke 正常返回 → 放行,后续逻辑不变
每个协程独立重试,失败后 sleep 200ms 再试,直到成功。
不再有全局门闸,互不影响。
"""
import asyncio
@@ -26,11 +25,6 @@ _THINK_RE = re.compile(r"<think>.*?</think>", re.DOTALL | re.IGNORECASE)
# 匹配 ```json ... ``` 代码块,提取中间的 JSON 内容
_CODE_BLOCK_RE = re.compile(r"```(?:json\w*)?\s*\n?(.*?)\n?\s*```", re.DOTALL | re.IGNORECASE)
# ──────────── 健康门闸 ────────────
_gate_event = asyncio.Event() # clear=故障阻塞中, set=健康放行
_gate_event.set() # 初始状态:健康
_gate_probing = False # 是否已有探针在重试
def parse_llm_json(text: str) -> Any:
"""解析 AI 输出的 JSON,自动去除思考标签、markdown 代码块,容错处理"""
@@ -45,13 +39,11 @@ def parse_llm_json(text: str) -> Any:
async def ai_chat(llm: BaseChatModel, system_prompt: str, user_message: str, scene: Optional[str] = None) -> str:
"""异步调用 LLM,返回原始文本。接口异常时触发门闸阻塞重试直到成功。"""
global _gate_probing
"""异步调用 LLM,返回原始文本。失败后独立重试直到成功。"""
attempt = 0
while True:
# 门闸检查:如果当前处于故障状态,阻塞等待直到探针恢复门闸
await _gate_event.wait()
attempt += 1
try:
messages = [
SystemMessage(content=system_prompt),
@@ -64,52 +56,8 @@ async def ai_chat(llm: BaseChatModel, system_prompt: str, user_message: str, sce
await _save_call_log(scene, system_prompt, user_message, content)
return content
except Exception as e:
# 接口失败 → 关闭门闸
if _gate_event.is_set():
_gate_event.clear()
log.error("AI 接口异常,门闸关闭,全部协程阻塞等待恢复: {}", e)
if not _gate_probing:
# 当前协程成为探针,负责退避重试直到成功
_gate_probing = True
await _probe_until_recover(llm, system_prompt, user_message)
# 探针恢复后,回到 while 循环顶部用业务数据重新调用
continue
# 已有探针在工作,回到 while 顶部 await _gate_event.wait() 等待唤醒
async def _probe_until_recover(llm: BaseChatModel, system_prompt: str, user_message: str) -> str:
"""探针:用轻量 hello 消息探测接口是否恢复,成功后打开门闸,原协程重新走正常流程"""
global _gate_probing
delay = 0.2 # 探测间隔 200ms
attempt = 0
try:
while True:
await asyncio.sleep(delay)
attempt += 1
try:
log.info("探针第{}次探测开始...", attempt)
messages = [
SystemMessage(content="You are a helpful assistant."),
HumanMessage(content="hello"),
]
await llm.ainvoke(messages)
# 成功 → 打开门闸,唤醒所有等待协程
_gate_probing = False
_gate_event.set()
log.info("探针第{}次探测成功,AI 接口恢复,门闸打开", attempt)
break
except Exception as e:
log.warning("探针第{}次探测失败: {}", attempt, e)
except (asyncio.CancelledError, Exception) as e:
# 探针意外死亡 → 必须恢复门闸状态,否则所有 worker 永远阻塞
_gate_probing = False
_gate_event.set()
log.error("探针协程意外退出,强制打开门闸: {}", e)
return ""
log.warning("AI 调用失败(第{}次), 200ms 后重试: {}", attempt, e)
await asyncio.sleep(0.2)
async def ai_chat_json(llm: BaseChatModel, system_prompt: str, user_message: str, scene: Optional[str] = None) -> Any: