do you want to pollute the context with blurbs for docs in disparate topics? cascade filtering, even with naïve bm25, helps reduce the amount of _noise_ that's pushed into the context window. if we reduce the amount of results to consider, further filtering or reranking, with more expensive options, becomes realistic. one could even put a cheaper model in front to further clean the results.