Does the LLM auto-tagging and embedding pipeline run on the device, or are they remote calls?