Update app.py

app.py

@@ -127,13 +127,14 @@ def stream_chat(message: str, history: list, model: str, temperature: float, max
                 "keep_alive": "60s",
             },
         )
-            
+
         buffer = ""
         for chunk in response:
             buffer += chunk["message"]["content"]
             yield buffer
 
 
+
 def main(message: str, history: list, model: str, temperature: float, max_new_tokens: int, top_p: float, top_k: int, penalty: float):
     if message.startswith("/"):
         resp = ollama_func(message)
@@ -144,7 +145,7 @@ def main(message: str, history: list, model: str, temperature: float, max_new_to
         else:
             if not process:
                 launch()
-            stream_chat(
+            answer = stream_chat(
                 message,
                 history,
                 model,
@@ -154,6 +155,7 @@ def main(message: str, history: list, model: str, temperature: float, max_new_to
                 top_k,
                 penalty             
             )
+            yield answer
 
 
 chatbot = gr.Chatbot(height=600, placeholder=DESCRIPTION)