Sparse retrieval is the approach of finding relevant documents using sparse, keyword-based representations such as BM25 or TF-IDF, where each document is represented by a vector over the entire vocabulary with most entries zero. It is the classic form of search that powered information retrieval long before dense embeddings arrived.
The word sparse reflects that these representations have a value only for the words actually present, leaving the vast majority of vocabulary dimensions at zero. Matching is based on shared terms and their importance, which makes sparse retrieval excellent at exact matching: product codes, names, error messages, and rare jargon that must be matched precisely and that dense embeddings may not represent reliably.
Sparse retrieval’s strength — exactness — is also its limit, since it cannot match synonyms or paraphrases that share no words. This is the mirror image of dense retrieval, which captures meaning but can miss exact terms. The two are complementary, which is why hybrid search combines them: sparse retrieval supplies lexical precision while dense retrieval supplies semantic breadth, and fusing their results captures both.