I think this is a really underexplored use case for LLMs. LLM embeddings are really good at encoding rich semantic information that’s easy to query and hack around with in a variety of ways. Retrieving primary sources that correspond to one or many thematic dimensions is one such case for embeddings, but most applications that do this portray it as a driver of RAG chatbots, when it could be an end in and of itself.

I have an app that does your book recommending idea but with Wikipedia articles. I am trying to release it soon, once I get past my perfectionism, if anyone is interested. Expanding to non Wikipedia sources is an eventual goal.

I basically never want to read chatbot output for pleasure. I want to read primary sources.

On one hand, that's a good idea.

On the other hand, do you really want to get stuck in a bubble and only read the same thing that you've read before and nothing new?

If it's bad when Facebook does it, it's bad with book recommendations too if you ask me.

And on the gripping hand...