Update app.py

app.py

@@ -18,6 +18,7 @@ def respond(message, history: list[tuple[str, str]]):
     messages.append({"role": "user", "content": message})
 
     response = ""
+    previous_response = ""
     for token_message in client.chat_completion(
         messages,
         max_tokens=max_tokens,
@@ -25,12 +26,18 @@ def respond(message, history: list[tuple[str, str]]):
         temperature=temperature,
         top_p=top_p,
     ):
-        # Retrieve token content; if empty, break the loop
         token = token_message.choices[0].delta.get("content", "")
         if not token:
             break
         response += token
-        yield response
+        # Only yield if new content was added
+        if response != previous_response:
+            yield response
+            previous_response = response
+
+        # Optional: break out if the response is too long to avoid infinite loops
+        if len(response) > 3000:  # adjust threshold as needed
+            break
 
 demo = gr.ChatInterface(respond)