The idea, in plain words
Meaning and term signals
A dense embedding spreads information across many learned coordinates. It can help match a paraphrase even when the wording differs.
A sparse representation has mostly zero values. BM25 is a classical lexical ranking method based on term statistics; learned sparse encoders can also represent semantic expansions. Sparse does not always mean ‘exact words only’.
Represent the question and passage in the dense model's compatible space.
Compare all dense dimensions, rather than a displayed preview.
Inspect which query terms contribute to the BM25 score.
Compare an exact product-code question with a paraphrase.