Sparse Retrieval
Also called Lexical Retrieval
Finding documents by the actual words they share with your query — old-school, but precise.
Think of it like
Ctrl-F across every document, weighted so rare words count more than common ones.
Example
Searching an exact error code "ERR_0x8007" returns the one doc that literally contains it, no semantics required.
How it actually works
Sparse retrieval represents text as high-dimensional word-count vectors (mostly zeros, hence "sparse") and scores overlap, classically with BM25. It nails exact terms, rare identifiers, and jargon, but it is literal — it misses synonyms and paraphrase entirely. Modern systems keep it around precisely because dense embeddings are bad at the things sparse is great at.
For product teams
Reliable for precise, keyword-heavy queries; blind to meaning, so pair it with dense for best coverage.
For engineers
Bag-of-words term weighting (TF-IDF/BM25) over an inverted index; exact-match strong, semantically blind.
Related
- BM25 — BM25 is the classic scoring method.
- Dense Retrieval — Contrasts with meaning-based dense retrieval.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome