Update app.py

app.py

@@ -27,13 +27,19 @@ def respond(
 
     response = ""
 
-    for message in client.chat_completion(
-        messages,
-        max_tokens=max_tokens,
-        stream=True,
-        temperature=temperature,
-        top_p=top_p,
-    ):
+response = client.text_generation(
+    prompt=message,  # Only pass the user message
+    max_new_tokens=max_tokens,
+    temperature=temperature,
+    top_p=top_p,
+    stream=True,
+)
+
+generated_text = ""
+for chunk in response:
+    generated_text += chunk
+    yield generated_text
+
         token = message.choices[0].delta.content
 
         response += token