That defeats the purpose then.

If the goal is to say "hi, my local LLM, recommend me a new movie that I like, because god forbid I'll take a risk and grow", then you having to curate the data feed going into LLM to describe new movies is in practice this:

• you'll have to find sources of data (websites, threads, communicties)
• verify each and everyone for poisoning BEFORE adding to your deterministic vectorization (see what Netflix did for recommendation algorithm recently) and applying custom LLM-like in architecture model.

Tl;dr;

Either you can use LLM and will be steered towards greatest quarterly revenue OR you have to put the same effort as if reading all reviews from multiple sources by yourself

And in both cases you still end up with the same outcome "maybe you'll like it" :P