Update app.py

app.py

@@ -13,9 +13,9 @@ DB_FAISS_PATH = 'vectorstore/db_faiss'
 def load_llm():
     # Load the locally downloaded model here
     llm = CTransformers(
-        model = "https://huggingface.co/TheBloke/Llama-2-7B-Chat-GGML/blob/main/llama-2-7b-chat.ggmlv3.q8_0.bin",
+        model = "llama-2-7b-chat.ggmlv3.q2_K.bin",
         model_type="llama",
-        max_new_tokens = 256,
+        max_new_tokens = 512,
         temperature = 0.2
     )
     return llm