If there are many common services for which you can precompute the embeddings then with a little record keeping and analysis you could figure out some likely questions or requests and pregenerate the responses. That way you could just use similarity search on the question or command you say and skip using the LLM. It would be interesting to try using the LLM to predict some of these based on information available ahead of time like calendar events, weather, recent prompt history, recently played media, today’s headlines, recent browser history, etc. It’d be your own recommendation algorithm.