diff --git a/src/config/base.yaml b/src/config/base.yaml
index f7e34f54..9ff69532 100644
--- a/src/config/base.yaml
+++ b/src/config/base.yaml
@@ -4,9 +4,10 @@ name: base
 ## model
 ### model_provider, option in deepseek, zhipu
 model_provider: zhipu
+model_name: null # for default
 
 ## model dir 可以写相对路径和绝对路径
-### 相对路径是相对于环境变量中 MODEL_ROOT_DIR 的路径
+### 相对路径是相对于环境变量 (.env) 中 MODEL_ROOT_DIR 的路径
 model_local_paths:
   bge-large-zh-v1.5: bge-large-zh-v1.5
   oneke: oneke
\ No newline at end of file
diff --git a/src/models/README.md b/src/models/README.md
new file mode 100644
index 00000000..732f26bd
--- /dev/null
+++ b/src/models/README.md
@@ -0,0 +1,42 @@
+## 模型说明
+
+### 1. 对话模型支持
+
+模型仅支持通过API调用的模型，如果是需要运行本地模型，则建议使用 vllm 转成 API 之后使用。
+
+|模型供应商(`config.model_provider`)|默认模型(`model.model_name`)|配置项目(`.env`)|
+|:-|:-|:-|
+|`qianfan`|`ernie_speed`|`QIANFAN_ACCESS_KEY`, `QIANFAN_SECRET_KEY`|
+|`zhipu`|`glm-4`|`ZHIPUAPI`|
+|`deepseek`|`deepseek-chat`|`DEEPSEEKAPI`|
+|`vllm`|`vllm`|`VLLM_API_KEY`, `VLLM_API_BASE`|
+
+vllm 部署参考脚本：
+
+```bash
+python -m vllm.entrypoints.openai.api_server --model ~/models/Meta-Llama-3-8B-Instruct --served-model-name vllm --trust-remote-code
+```
+
+*openai 没条件测，不知道
+
+### 2. 向量模型支持
+
+
+|模型名称(`config.embed_model`)|默认路径|可配置项目（`config.model_local_paths`）|
+|:-|:-|:-|
+|`bge-large-zh-v1.5`|`BAAI/bge-large-zh-v1.5`|`bge-large-zh-v1.5`|
+
+### 3. 重排序模型支持
+
+
+例如：
+
+```yaml
+model_provider: qianfan
+model_name: null # for default
+
+## model dir 可以写相对路径和绝对路径
+### 相对路径是相对于环境变量 (.env) 中 MODEL_ROOT_DIR 的路径
+model_local_paths:
+  bge-large-zh-v1.5: bge-large-zh-v1.5
+```
diff --git a/src/models/__init__.py b/src/models/__init__.py
index bae3a78e..35cd0c5d 100644
--- a/src/models/__init__.py
+++ b/src/models/__init__.py
@@ -20,6 +20,10 @@ def select_model(config):
         from models.chat_model import Qianfan
         return Qianfan(model_name)
 
+    elif model_provider == "vllm":
+        from models.chat_model import VLLM
+        return VLLM(model_name)
+
     elif model_provider is None:
         raise ValueError("Model provider not specified, please modify `model_provider` in `src/config/base.yaml`")
     else:
diff --git a/src/models/chat_model.py b/src/models/chat_model.py
index f6b8a8fc..94e4f7f0 100644
--- a/src/models/chat_model.py
+++ b/src/models/chat_model.py
@@ -56,6 +56,13 @@ class Zhipu(OpenAIBase):
         base_url = "https://open.bigmodel.cn/api/paas/v4/"
         super().__init__(api_key=api_key, base_url=base_url, model_name=model_name)
 
+class VLLM(OpenAIBase):
+    def __init__(self, model_name=None):
+        model_name = model_name or "vllm"
+        api_key = os.getenv("VLLM_API_KEY")
+        base_url = os.getenv("VLLM_API_BASE")
+        super().__init__(api_key=api_key, base_url=base_url, model_name=model_name)
+
 
 import qianfan
 
@@ -99,4 +106,6 @@ class Qianfan:
             messages=messages,
             stream=False,
         )
-        return QianfanResponse(response["body"]["result"])
\ No newline at end of file
+        return QianfanResponse(response["body"]["result"])
+
+
diff --git a/src/views/common_view.py b/src/views/common_view.py
index 8e0b70dd..0436ead6 100644
--- a/src/views/common_view.py
+++ b/src/views/common_view.py
@@ -44,13 +44,14 @@ def chat():
     def generate_response():
         content = ""
         for delta in model.predict(messages, stream=True):
-            content += delta.content
-            response_chunk = json.dumps({
-                "history": history_manager.update_ai(content),
-                "response": content,
-                "refs": refs  # TODO: 优化 refs，不需要每次都返回
-            }, ensure_ascii=False).encode('utf8') + b'\n'
-            yield response_chunk
+            if delta.content:
+                content += delta.content
+                response_chunk = json.dumps({
+                    "history": history_manager.update_ai(content),
+                    "response": content,
+                    "refs": refs  # TODO: 优化 refs，不需要每次都返回
+                }, ensure_ascii=False).encode('utf8') + b'\n'
+                yield response_chunk
 
     return Response(generate_response(), content_type='application/json', status=200)