diff --git a/backend/app/main.py b/backend/app/main.py
index 52be5308..66de186a 100644
--- a/backend/app/main.py
+++ b/backend/app/main.py
@@ -5,8 +5,10 @@ from sqlalchemy.future import select
 from typing import List
 from pydantic import BaseModel
 from enum import Enum
+from pathlib import Path
 import uuid
 import asyncio
+import os
 
 from .database import get_db, engine, Base
 from .models import ContextNode, Experiment, Signal, SignalFeedback
diff --git a/backend/app/services/llm_client.py b/backend/app/services/llm_client.py
index 6fbf18aa..72aa9e81 100644
--- a/backend/app/services/llm_client.py
+++ b/backend/app/services/llm_client.py
@@ -123,14 +123,24 @@ class OpenAICompatibleClient(LLMProvider):
     """OpenAI 兼容客户端 (适用于 DeepSeek, Qwen 等)"""
 
     def __init__(self, api_key: str, base_url: str, model: str):
-        if not api_key:
-            raise ValueError(f"请设置 {base_url} 的 API_KEY")
         self.model = model
-        self.client = httpx.Client(
-            base_url=base_url,
-            timeout=120.0,
-            headers={"Authorization": f"Bearer {api_key}"}
-        )
+        self.base_url = base_url
+        self.api_key = api_key
+        # 延迟初始化客户端，在实际调用时才检查 API_KEY
+        self._client = None
+
+    @property
+    def client(self):
+        """延迟初始化客户端"""
+        if self._client is None:
+            if not self.api_key:
+                raise ValueError(f"请设置 {self.base_url} 的 API_KEY")
+            self._client = httpx.Client(
+                base_url=self.base_url,
+                timeout=120.0,
+                headers={"Authorization": f"Bearer {self.api_key}"}
+            )
+        return self._client
 
     def generate(self, prompt: str, **kwargs) -> str:
         response = self.client.post("/chat/completions", json={
@@ -154,7 +164,7 @@ class OpenAICompatibleClient(LLMProvider):
 
     def get_embedding(self, text: str) -> List[float]:
         # DeepSeek 不支持 embedding，使用替代方案或返回空
-        if "deepseek" in self.client.base_url:
+        if "deepseek" in self.base_url:
             print("[WARNING] DeepSeek 不支持 embedding API，使用简化向量")
             # 简化方案：基于文本内容生成伪向量（实际使用时应该用其他服务）
             import hashlib