Only extremely basic requests are processed on device, like “what time is it”. Anything else is still processed on Apple’s servers.
This should bring a 10x magnitude of improvement and justify hardware upgrades to satiate the RAM and coprocessor thirst of LLMs.
Spotlight does pretty extensive if “traditional” index-based search of your local personal data, along with certain privacy-preserving web queries. You can imagine what on-device LLM-powered summaries would bring to this type of search and suggested actions.