Sparse Retrieval
Classic keyword retrieval (like BM25) scoring documents by term overlap — sparse because most vocabulary entries are zero.
Sparse retrieval is keyword-based search — methods like BM25 and TF-IDF — that score documents by the overlap of query terms, represented as high-dimensional but mostly-zero (sparse) vectors over the vocabulary. It is exact, interpretable, and needs no training.
Worked example: searching “error code E4021” with BM25 nails the document containing that exact string, weighting rare terms highly — where a dense embedding might drift to semantically similar but wrong pages. Gotcha: sparse retrieval cannot match synonyms or paraphrases (“car” will not find “automobile”), which is exactly where dense retrieval helps — modern hybrid search runs both and fuses the scores to get the strengths of each.