Update app.py

app.py

@@ -220,12 +220,13 @@ Context:
             #     )
             # )
             
-            return self.llm.chat.completions.create(
+            for x in self.llm.chat.completions.create(
                         model=model_name, 
                     	messages=messages, 
                     	max_tokens=500,
                         stream=True
-                        )
+                        ):
+                yield x
 
         except Exception as e:
             error_msg = f"Error during query processing: {str(e)}"
@@ -252,8 +253,7 @@ def create_gradio_interface(rag_system: AdvancedRAGSystem) -> gr.Blocks:
     def query_fin(question):
         """Query system and update history with error handling"""
         try:
-            for x in rag_system.query(question):
-                yield x.choices[0].delta.content
+            yield rag_system.query(question).choices[0].delta.content
         except Exception as e:
             pass
             
@@ -317,7 +317,9 @@ def create_gradio_interface(rag_system: AdvancedRAGSystem) -> gr.Blocks:
         query_button.click(
             fn=query_fin,
             inputs=[question_input],
-            outputs=[answer_output]
+            outputs=[answer_output],
+            api_name="stream_response",
+            show_progress=False
         )
 
         query_button.click(