Dense Retrieval¶

Dense retrieval methods represent documents and queries as fixed-size dense vectors (embeddings) and rank candidates based on similarity in embedding space. Unlike sparse retrieval methods (e.g., BM25), dense retrievers learn to embed semantically similar content near each other, capturing query-document relevance through continuous representations.

Dense retrievers typically use dual-encoder architectures where queries and documents are encoded independently. This enables efficient batch retrieval operations and supports both retrieval and generation in unified frameworks.

Key papers¶

Atlas: Few-shot Learning with Retrieval Augmented Language Models — Uses Contriever, a dense retriever based on continuous embeddings, within a retrieval-augmented language model

Information Retrieval (general retrieval)
Retrieval-Augmented Generation (integration with generation)
Neural networks (technical foundation)

Dense Retrieval¶

Key papers¶

Related topics¶