Update app.py

app.py

@@ -22,13 +22,14 @@ def runpod_chat(question, history=None):
         stream=True,
     )
     
-    # Stream the response and add to history
-    full_response = ""
+    # Stream the response and accumulate full response before displaying
+    full_response = "RunPod: "
     for message in response_stream:
-        response = message.choices[0].delta.content if message.choices[0].delta.content is not None else "No response."
-        full_response += "RunPod: " + response + "\n"
-        history.append({"role": "assistant", "content": response})
-        time.sleep(0.3)  # Simulate typing delay
+        part = message.choices[0].delta.content if message.choices[0].delta.content is not None else ""
+        full_response += part
+
+    # Append the full response to history once complete
+    history.append({"role": "assistant", "content": full_response})
 
     return full_response, history  # Return full response and updated history to maintain state