I’m sure it’s RAG at best. There’s no way I can conceive of that they’re actually training individual models for each user in a performant or economical fashion.

More likely, as you said, they’re just zero-shotting the relevant personal data into the context window. And honestly, I’d be a little surprised if they had a smaller model trying to evaluate relevance; a simple heuristic or basic frequency analysis algorithm would probably perform about as well and be a lot cheaper. The big final model can probably toss away the noise well enough.