Dense Retrieval

Dense retrieval is an information retrieval approach where both queries and documents are encoded as dense (mostly non-zero) vector embeddings using neural models, and relevance is determined by vector similarity (typically cosine similarity or dot product). This contrasts with sparse retrieval methods like BM25 that use high-dimensional sparse vectors based on term frequencies.

The key innovation of dense retrieval is that it captures semantic similarity rather than lexical overlap. A query about "automobile maintenance" can match a document about "car repair" because the neural encoder maps semantically related text to nearby points in embedding space. Dense Passage Retrieval (DPR) is a foundational dense retrieval method that uses separate BERT encoders for queries and documents, trained with contrastive learning on positive and negative query-document pairs.

Dense retrieval requires a vector database for efficient similarity search and GPU resources for encoding. While it excels at semantic matching, it can miss exact-match queries for rare terms, codes, and names. This limitation is why production systems combine dense retrieval with sparse retrieval (BM25) in hybrid search configurations.