Dense Retrieval
Dense retrieval finds relevant passages by comparing the meaning-based vector embeddings of a query and your content, matching on semantics rather than exact words.
Dense retrieval matches by meaning, using embeddings. It converts the query and every candidate passage into dense vector embeddings and returns the passages whose vectors sit closest in meaning — the engine behind vector search. "Dense" refers to these rich numeric representations, as opposed to the word-count vectors of sparse retrieval.
Its rise is why writing for concepts beats writing for keywords. Dense retrieval can match a query to your passage even with no shared words, as long as the meaning lines up — so a single, clearly-expressed idea per passage, the extractability discipline, is what lands you a precise semantic match. Vague or multi-topic passages embed into fuzzy space and match nothing strongly.
Example. Dense retrieval can surface a page about "lowering cholesterol through diet" for the query "foods that reduce LDL," because the meanings align — a match keyword-only methods would miss.
Relevant pillar
Related terms
- Vector SearchVector search is a retrieval method that finds passages by meaning rather than keywords, comparing the numeric embedding of a query against the embeddings of indexed content to surface the closest matches.
- EmbeddingsEmbeddings are numerical representations of text that capture its meaning, letting AI systems find passages that are semantically related to a query even when they share no exact keywords.
- Sparse RetrievalSparse retrieval finds relevant content by matching actual words and their importance, using classic methods like BM25, and still complements meaning-based retrieval in modern systems.