Comment by m0wer

Probably what you want is not an LLM but just the embeddings for clustering. It's much lighter and would work well with new material as well.

I've tested it out for filtering RSS feeds and has worked pretty well [1].

[1] https://github.com/m0wer/rssfilter