feat:修复投顾agent功能

tool/llm.py

@@ -103,6 +103,21 @@ class LLMClient:
     def __init__(self, cfg: LLMCfg | None = None):
         self.cfg = cfg or settings.llm
         self.backends = resolve_backends(self.cfg)
+        self._clients: dict[str, httpx.AsyncClient] = {}
+
+    def _client_for(self, backend: Backend) -> httpx.AsyncClient:
+        clients = getattr(self, "_clients", None)
+        if clients is None:
+            clients = {}
+            self._clients = clients
+        if backend.name not in clients:
+            clients[backend.name] = backend.client(self.cfg.timeout)
+        return clients[backend.name]
+
+    async def aclose(self) -> None:
+        for client in getattr(self, "_clients", {}).values():
+            await client.aclose()
+        getattr(self, "_clients", {}).clear()
 
     # ---- 首选后端（健康检查/日志/embed 使用） -----------------------------
     @property
@@ -186,9 +201,9 @@ class LLMClient:
                 "stream": False,
             }
             try:
-                async with backend.client(self.cfg.timeout) as client:
-                    r = await client.post(url, headers=backend.headers, json=payload)
-                    r.raise_for_status()
+                client = self._client_for(backend)
+                r = await client.post(url, headers=backend.headers, json=payload)
+                r.raise_for_status()
                 choice = r.json()["choices"][0]
                 message = choice["message"]
                 content = message.get("content")
@@ -250,20 +265,16 @@ class LLMClient:
         response = None
         max_retries = max(1, int(getattr(self.cfg, "max_retries", 1)))
         retry_backoff_sec = float(getattr(self.cfg, "retry_backoff_sec", 0))
-        async with backend.client(self.cfg.timeout) as client:
-            for attempt in range(max_retries):
-                try:
-                    response = await client.post(
-                        url,
-                        headers=backend.headers,
-                        json=payload,
-                    )
-                    response.raise_for_status()
-                    break
-                except Exception:
-                    if attempt == max_retries - 1:
-                        raise
-                    await asyncio.sleep(retry_backoff_sec * (2**attempt))
+        client = self._client_for(backend)
+        for attempt in range(max_retries):
+            try:
+                response = await client.post(url, headers=backend.headers, json=payload)
+                response.raise_for_status()
+                break
+            except Exception:
+                if attempt == max_retries - 1:
+                    raise
+                await asyncio.sleep(retry_backoff_sec * (2**attempt))
         if response is None:  # pragma: no cover - defensive guard
             raise LLMFailError("Embedding 请求未返回响应")
         data = response.json()
@@ -286,10 +297,10 @@ class LLMClient:
             url = backend.base_url.removesuffix("/v1") + "/api/tags"
         else:
             url = f"{backend.base_url}/models"
-        async with backend.client(min(self.cfg.timeout, 10)) as client:
-            r = await client.get(url, headers=backend.headers)
-            if r.status_code >= 500:
-                r.raise_for_status()
+        client = self._client_for(backend)
+        r = await client.get(url, headers=backend.headers)
+        if r.status_code >= 500:
+            r.raise_for_status()
 
 
 # 全局单例：Agent 统一引入