Update app.py

app.py

@@ -139,7 +139,16 @@ def respond(
     prompt = format_prompt(message, history)
 
     # Generate response using the InferenceClient
-    response = client.generate(prompt, max_tokens=max_tokens, temperature=temperature, top_p=top_p)
+    response = ""
+    for message in client.chat_completion(
+        prompt,
+        max_tokens=max_tokens,
+        temperature=temperature,
+        top_p=top_p,
+    ):
+        token = message.choices[0].delta.content
+        response += token
+        yield response
 
     # Append the response to history
     history.append((message, response))
@@ -147,8 +156,6 @@ def respond(
     # Handle the selected task
     handle_input(task, storage_location, url1, url2, scrape_interval, content_type)
 
-    return response
-
 # Create Gradio interface
 demo = gr.ChatInterface(
     respond,