thought

BM25 Sparse Search

Sparse search is based on exact keyword matching (bag of words). BM25 improves TF-IDF with two additions: term frequency saturation (each subsequent occurrence provides a smaller gain, parameter k1) and document length normalization (parameter b). Rare words carry more weight (IDF). Strengths: exact names, technical codes, terms; weakness: does not understand synonyms ("kitty” will not find “cat"). Learnable variants (SPLADE, BGE-M3 sparse) add neural network weights to terms.