feat:修改nl2sql功能

tool/llm.py

@@ -25,6 +25,7 @@ logger = logging.getLogger("tool.llm")
 FALLBACK_REPLY = "抱歉，服务暂时不可用，请稍后再试，或拨打客服热线 400-XXX-XXXX。"
 
 
+
 @dataclass(frozen=True)
 class Backend:
     """一个 OpenAI 兼容端点及其模型；name 为 ollama / api。"""
@@ -34,6 +35,7 @@ class Backend:
     chat_model: str
     embed_model: str
     api_key: str = ""
+    api_trust_env: bool = True
 
     @property
     def is_ollama(self) -> bool:
@@ -50,7 +52,8 @@ class Backend:
         """ollama 走本机端点，必须绕过系统代理：httpx 会读取 Windows 系统代理，
         但不识别其 ProxyOverride 白名单，localhost 请求会被转发到代理并返回 502。
         API 后端保留 trust_env，远端接口可能依赖代理。"""
-        return httpx.AsyncClient(timeout=timeout, trust_env=not self.is_ollama)
+
+        return httpx.AsyncClient(timeout=timeout, trust_env=self.api_trust_env if not self.is_ollama else False)
 
 
 def resolve_backends(cfg: LLMCfg) -> list[Backend]:
@@ -61,6 +64,7 @@ def resolve_backends(cfg: LLMCfg) -> list[Backend]:
             base_url=cfg.ollama_base.rstrip("/") + "/v1",
             chat_model=cfg.ollama_chat_model,
             embed_model=cfg.ollama_embed_model,
+            api_trust_env=False,
         )
         if cfg.ollama_base and cfg.ollama_chat_model
         else None
@@ -72,6 +76,7 @@ def resolve_backends(cfg: LLMCfg) -> list[Backend]:
             chat_model=cfg.api_chat_model,
             embed_model=cfg.api_embed_model,
             api_key=cfg.api_key,
+            api_trust_env=cfg.api_trust_env,
         )
         if cfg.api_key and cfg.api_base and cfg.api_chat_model
         else None
@@ -224,8 +229,8 @@ class LLMClient:
         ) or backend.embed_model
         if not embed_model:
             raise LLMFailError(f"backend={backend.name} 未配置 embed 模型")
-        is_dashscope_compatible = "/compatible-mode/" in backend.base_url.lower()
-        if is_dashscope_compatible:
+        is_dashscope_multimodal = "vision" in embed_model.lower() or "multimodal" in embed_model.lower()
+        if is_dashscope_multimodal:
             base_url = backend.base_url.rstrip("/")
             marker = "/compatible-mode/v1"
             if base_url.lower().endswith(marker):
@@ -242,11 +247,27 @@ class LLMClient:
                 "input": texts,
                 "dimensions": self.cfg.embed_dimensions,
             }
+        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:
-            r = await client.post(url, headers=backend.headers, json=payload)
-            r.raise_for_status()
-        data = r.json()
-        if is_dashscope_compatible:
+            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()
+        if is_dashscope_multimodal:
             embeddings = data["output"]["embeddings"]
             return [
                 item["embedding"]