BM25
A classical information-retrieval ranking function that combines query-term frequency, document-length normalization, and inverse document frequency.
BM25 does not reward repeated occurrences of a term without limit; the term-frequency contribution approaches saturation. Length normalization also prevents long documents from gaining score merely because they contain more words.
Unlike dense embedding search, BM25 is a lexical matching signal. It remains useful when exact technical terms, identifiers, or error strings matter, and hybrid systems can combine it with vector retrieval. FAISS vector-search index architecture covers the dense side of that comparison.