diff --git a/vllm/entrypoints/openai/protocol.py b/vllm/entrypoints/openai/protocol.py
index e85e7e2b1ede9..97cfd797587c4 100644
--- a/vllm/entrypoints/openai/protocol.py
+++ b/vllm/entrypoints/openai/protocol.py
@@ -57,7 +57,7 @@ class UsageInfo(BaseModel):
 
 class ChatCompletionRequest(BaseModel):
     model: str
-    messages: Union[str, List[Dict[str, str]]]
+    messages: List[Dict[str, str]]
     temperature: Optional[float] = 0.7
     top_p: Optional[float] = 1.0
     n: Optional[int] = 1
diff --git a/vllm/entrypoints/openai/serving_chat.py b/vllm/entrypoints/openai/serving_chat.py
index 5635ac6c9e106..e5ae39e110a40 100644
--- a/vllm/entrypoints/openai/serving_chat.py
+++ b/vllm/entrypoints/openai/serving_chat.py
@@ -80,7 +80,7 @@ def get_chat_request_role(self, request: ChatCompletionRequest) -> str:
         if request.add_generation_prompt:
             return self.response_role
         else:
-            return request.messages[-1].role
+            return request.messages[-1]["role"]
 
     async def chat_completion_stream_generator(
             self, request: ChatCompletionRequest,