Fixed repetitive output

app.py

@@ -34,11 +34,12 @@ def convert_history(message, history):
 
 def ask(message, history):
     chat_history = convert_history(message, history)
-    chunks = llm.create_chat_completion(
-        messages = chat_history,
-        temperature = 0,
+    chunks = llm(
+        chat_history,
+        temperature = 0.2,
+        top_p=0.9,
         stream = True,
-        repeat_penalty = 1.2,
+        repeat_penalty = 1.05,
         max_tokens = 128,
     )