Dense Retrieval
Authoritative Definition
A semantic search paradigm utilizing dense neural embeddings, where all vector dimensions contain continuous real numbers representing the latent meaning of text.
Overview & Technical Description
Dense retrieval maps text—ranging from short queries to long documents—into a high-dimensional continuous vector space using transformer-based encoder models. Because these vectors capture semantic concepts rather than exact keywords, queries can successfully retrieve documents that use entirely different vocabulary but share the same underlying meaning. This approach contrasts sharply with traditional sparse retrieval (like BM25), which relies on term frequency and inverted indexes. In dense retrieval, relevance is determined by calculating spatial distances (such as cosine similarity or dot product) between the query vector and document vectors stored in a vector database. It has become the foundational retrieval layer for modern RAG applications due to its ability to understand context and intent.
Editorial Notes
Powered by transformer encoder models like BGE, E5, OpenAI text-embedding-3, and Voyage. Best results in production are usually achieved by combining dense retrieval with sparse retrieval in a hybrid search configuration.