Dense retrieval
The standard modern retrieval approach where queries and documents are encoded as dense embedding vectors and matched by similarity — distinct from sparse retrieval (BM25, keyword search).
Dense retrieval is what most people mean when they say "vector search" or "semantic search." The "dense" refers to the fact that every dimension of the embedding vector has a meaningful value (vs sparse retrieval where most dimensions are zero in techniques like BM25).
For modern RAG, dense retrieval is the workhorse — fast at scale, captures meaning, handles synonyms gracefully. Combined with sparse retrieval (BM25) in [hybrid search](/glossary/hybrid-search), it powers the most accurate production RAG systems.
In 2026 the choice between dense and hybrid is a maturity progression. Start with dense for simplicity; move to hybrid once you hit exact-match or proper-noun blind spots.
Frequently asked
Dense vs sparse retrieval?+
Dense (embeddings) for meaning. Sparse (BM25, keyword) for exact matches. Production RAG should use both via hybrid search — pure dense misses proper nouns and codes; pure sparse misses synonyms and rephrasings.
How fast is dense retrieval at scale?+
Vector databases like Pinecone and Qdrant handle 100M+ vectors with sub-100ms p95 latency. Past that, additional infrastructure (sharding, quantization, hierarchical indexes) keeps performance acceptable.