diff --git a/scripts/init_kb.py b/scripts/init_kb.py
index 7aef308..b50ad4c 100644
--- a/scripts/init_kb.py
+++ b/scripts/init_kb.py
@@ -13,6 +13,31 @@ sys.path.insert(0, str(Path(__file__).parent.parent))
 
 load_dotenv()
 
+
+def download_embeddings_model():
+    """预下载 Qwen3-Embedding 模型（从 HuggingFace）。
+
+    用法: python scripts/init_kb.py --download-model
+    """
+    model_name = os.getenv("LOCAL_EMBED_MODEL", "Qwen/Qwen3-Embedding-0.6B")
+    print(f"正在下载嵌入模型: {model_name}")
+    print("如遇网络超时，可手动执行以下命令后重试:")
+    print(f"  huggingface-cli download {model_name} --local-dir ./models/{model_name.replace('/', '_')}")
+    print()
+
+    try:
+        from langchain_huggingface import HuggingFaceEmbeddings
+    except ImportError:
+        print("错误: 请先安装 huggingface 依赖")
+        print("  pip install langchain-huggingface sentence-transformers")
+        return
+
+    # HuggingFaceEmbeddings 会在首次调用时自动下载模型
+    embeddings = HuggingFaceEmbeddings(model_name=model_name)
+    # 调用一次以确保完全下载
+    embeddings.embed_query("测试")
+    print(f"嵌入模型下载完成: {model_name}")
+
 from backend.embeddings import get_embeddings
 
 
@@ -84,4 +109,12 @@ def main():
 
 
 if __name__ == '__main__':
-    main()
+    import argparse
+    parser = argparse.ArgumentParser(description='初始化 Chroma 知识库')
+    parser.add_argument('--download-model', action='store_true', help='仅下载嵌入模型到本地')
+    args = parser.parse_args()
+
+    if args.download_model:
+        download_embeddings_model()
+    else:
+        main()