What is Dense Retrieval

How It Works

Dense retrieval uses dual encoder models to independently map queries and documents into a shared embedding space. At query time, the query is encoded into a vector, and the most similar document vectors are retrieved using approximate nearest neighbor search. This captures semantic meaning rather than relying on exact keyword overlap.

Technical Details

Models like DPR, E5, and BGE use transformer-based dual encoders trained with contrastive learning on query-document pairs. Document embeddings are pre-computed and indexed in vector databases. Query latency is dominated by the ANN search step since query encoding is a single forward pass. Typical embedding dimensions range from 384 to 1024.

Best Practices

• Fine-tune the retrieval model on in-domain query-document pairs for best performance
• Use hard negative mining during training to improve discrimination between similar documents
• Combine dense retrieval with sparse retrieval (BM25) in a hybrid approach for optimal recall
• Pre-compute and cache document embeddings to minimize latency at query time

Common Pitfalls

• Relying solely on dense retrieval for queries that require exact keyword matching
• Not normalizing embeddings before cosine similarity computation
• Using a model trained on general data for a highly specialized domain without adaptation
• Underestimating the storage cost of dense embeddings at scale

Advanced Tips

• Implement late interaction models like ColBERT for token-level matching with efficient retrieval
• Use knowledge distillation from cross-encoder rerankers to improve bi-encoder quality
• Apply in-batch negatives during training to increase effective negative sample size
• Leverage multi-vector representations for complex queries that span multiple concepts