Update app.py

app.py

@@ -72,13 +72,15 @@ async def chat(gen:GenModel):
                 {"role": "assistant", "content": gen.system},
             ]
         st = time()
+        
+        messages.append({"role": "user", "content": gen.question})
         output = llm_chat.create_chat_completion(
             messages = messages,
             temperature=gen.temperature,
             seed=gen.seed,
             #stream=True
         )
-        messages.append({"role": "user", "content": gen.question})
+        #messages.append({"role": "user", "content": gen.question})
         print(output)
         et = time()
         output["time"] = et - st