Dense retrieval uses learned embedding vectors to find semantically relevant documents — encoding queries and documents into dense vector representations using bi-encoder models, then finding nearest neighbors in embedding space, enabling semantic search that understands meaning rather than relying on exact keyword matches.

How Dense Retrieval Works

• Bi-Encoder: Separate encoders for queries and documents produce independent embeddings.
• Indexing: Pre-compute document embeddings, store in vector database.
• Search: Encode query, find nearest document vectors via ANN search.
• Speed: Sub-millisecond search over millions of documents.

Advantages Over Sparse Retrieval (BM25)

• Semantic Understanding: "car" matches "automobile" and "vehicle."
• Zero-Shot: Works for unseen queries without keyword overlap.
• Multilingual: Cross-language retrieval with multilingual encoders.

Limitations: May miss exact keyword matches; hybrid (dense + sparse) retrieval often works best.

Dense retrieval powers modern RAG pipelines — enabling LLMs to find relevant context through semantic understanding rather than keyword matching.