Update vlm.py

vlm.py

@@ -97,7 +97,7 @@ def build_messages(message: dict, history: list[tuple]):
 # get response
 #
 def get_response(messages: list[dict]):
-    """Stream the model's response to the chat interface.
+    """Get the model's response.
     
     Args:
         messages: list of messages to send to the model
@@ -106,3 +106,16 @@ def get_response(messages: list[dict]):
     logger.info(f"{response=}")
     return response.choices[0].message.content
 
+#
+# stream response
+#
+def stream_response(messages: list[dict]):
+    """Stream the model's response.
+    
+    Args:
+        messages: list of messages to send to the model
+    """
+    response = ""
+    for chunk in client.chat.stream(model=model_id, messages=messages):
+        response += chunk.data.choices[0].delta.content
+        yield response