Update app.py

app.py

@@ -25,7 +25,7 @@ def respond(message, chat_history, system_message, max_tokens, temperature, top_
     messages.append({"role": "user", "content": message})
     
     try:
-        response = ""
+        full_response = ""
         for chunk in client.text_generation(
             prompt="\n".join([f"{m['role']}: {m['content']}" for m in messages]),
             max_new_tokens=max_tokens,
@@ -36,9 +36,11 @@ def respond(message, chat_history, system_message, max_tokens, temperature, top_
             if stop_event.is_set():
                 break
             if chunk:
-                response += chunk
-                chat_history.append((message, response))
-                yield chat_history
+                full_response += chunk
+                yield chat_history + [(message, full_response)]
+        
+        chat_history.append((message, full_response))
+        yield chat_history
     except Exception as e:
         error_message = f"오류 발생: {str(e)}"
         chat_history.append((message, error_message))