Sparse retrieval
Sparse retrieval represents queries and documents as vectors over a vocabulary where most entries are zero and each non-zero entry is a word weight, so search can use fast inverted indexes like keyword search.
Also called: sparse embeddings, SPLADE, learned sparse retrieval
By Divya Sharma · Checked Nov 15, 2026 · How we review
In a dense embedding, every value is non-zero and none has a fixed meaning. In a sparse vector, each position stands for a vocabulary term, and only the terms that matter for the text get a weight. BM25 is the classic sparse method, with weights from a fixed formula.
Learned sparse models
Models such as SPLADE use a neural network to set the weights. They can also add related terms the text never used, a step called expansion, so a document about "automobiles" gets weight on "car" too. The result keeps the strengths of keyword search, exact term matching and inverted indexes, while closing part of the gap with dense models on meaning.
Why use it
Sparse vectors are easy to inspect: you can see which words drove a match. They do well on names, codes and jargon that dense models blur. They also run on existing search engines. A common setup is hybrid search with one sparse and one dense retriever, merged before reranking.
Limits
Learned sparse models need a model call at indexing and query time, unlike BM25. Their vocabulary is fixed by the tokenizer, so they can struggle across languages. Support varies between vector databases; check yours before committing.
Related terms
- BM25
- BM25 is a ranking formula for keyword search. It scores each document by how often the query's words appear in it, giving more weight to rare words and less credit to long documents and repeated terms.
- Hybrid search
- Hybrid search runs a keyword search and a vector search on the same query and merges the two ranked lists, so results can match on exact words as well as on meaning.
- Embedding
- An embedding is a list of numbers (a vector) that represents a piece of text, an image or other data, produced so that items with similar meaning end up with vectors close together.
- Vector search
- Vector search finds items by comparing their embeddings to the embedding of a query and returning the closest ones, so results match on meaning rather than on shared keywords.
- Reranking
- Reranking is a second search step that takes a short list of candidate documents and reorders them by relevance to the query, using a more accurate but slower model than the first search.