Update main.py

main.py

@@ -62,7 +62,8 @@ def response_generator(query: str) -> str:
 
     # prepare HF text-generation LLM
     hf = HuggingFaceEndpoint(
-        endpoint_url=f"https://api-inference.huggingface.co/models/{model}",
+        # endpoint_url=f"https://api-inference.huggingface.co/models/{model}",
+        endpoint_url=f"https://router.huggingface.co/hf-inference/models/{model}",
         task="text-generation",
         huggingfacehub_api_token=hf_api_key,
         model_kwargs={