Update app.py

app.py

@@ -84,7 +84,9 @@ def get_chain(temperature):
     retriever = load_retriever(docstore_path,chroma_path,embeddings,child_splitter,parent_splitter)
     
     # Replace the local OLMOLLM with the Hugging Face model
-    
+    callback_handler.on_llm_end = lambda response, *args, **kwargs: None
+    # Callbacks support token-wise streaming
+    callback_manager = CallbackManager([callback_handler])
     llm = LlamaCpp(
         model_path=str(model_path),
         callback_manager=callback_manager,