Dense Retrieval

Dense retrieval is a neural approach to information retrieval that represents both queries and documents as dense vectors (embeddings) in a continuous vector space. Relevance is determined by the similarity between the query vector and document vectors, typically measured using cosine similarity or dot product. The term 'dense' contrasts with 'sparse' retrieval methods like BM25 and TF-IDF, where documents are represented as sparse vectors with most dimensions being zero (one dimension per vocabulary term). Dense vectors are comparatively compact (typically 256-1024 dimensions) with meaningful values in every dimension. Dense retrieval models are trained to produce similar embeddings for semantically related queries and documents, even when they share few or no common words. This is achieved through contrastive learning on pairs of queries and relevant passages, teaching the model to map related text to nearby points in the vector space.

Dense retrieval enables truly semantic search — finding relevant content based on meaning rather than keyword overlap. This is transformative for enterprise search and RAG systems where users often query using different terminology than what appears in the source documents. The practical benefit is significant. An employee searching for 'how to request time off' can find the relevant HR document even if it is titled 'Annual Leave Application Procedure' and never mentions 'time off'. This reduces search frustration and ensures that relevant information is discoverable. Dense retrieval is most effective when combined with sparse retrieval in a hybrid approach. Dense methods handle semantic similarity well but can miss exact-match queries (product codes, specific names). Hybrid search combines the strengths of both approaches for the best overall retrieval quality.