Hello, future.

app.py

@@ -35,9 +35,12 @@ lcpp_model = Llama(model_path=mdlpath)
 
 def m3b_talk(text):
     resp = ""
-    for token in lcpp_model("<s> [|User|]" + text + "</s> [|Assistant|]", stop=["[|User|]", "\n"], echo=True):
+    formattedQuery = "<s> [|User|]" + text + "</s> [|Assistant|]"
+
+    for token in lcpp_model(formattedQuery, stop=["[|User|]", "\n"], echo=True):
         resp += token
-    return resp
+    answer = resp["choices"][0]["text"].replace(formattedQuery, "")
+    return answer
 
 def main():
     logging.basicConfig(level=logging.INFO)