[ccarnino]

To guarantee that it does not illuminate and stays on track, I pass the relevant chapters in the context window, I pass the topics to cover, the ones already covered and the latest messages exchanged. So highly relevant content as input and a refined prompt do the trick.

[falcor84]

But context is only additive. You can give it the book content, but I don't think there's anything you can do to prevent it from bringing in unrelated info, unless you have a separate non-AI layer (e.g. connected to the RAG) that gives the user the particular snippets from the book, but even then you can't control what the AI bring to its interpretation of these snippets.

One thing that I think could be nice is if you could have it always keep the book in view, and just scroll to the appropriate place it is referring to, and then somehow highlight the specific snippet it is currently referring to, so that I could visually verify that it's giving me "grounded" information. Does your system do something like that?

[ccarnino]

It currently does not do that. It's a different way to "experience" a book. There's an argument where, like you proposed, book and tutor are on at the same time. I will keep it in mind