Glossary term
Glossary term
Memory and Retrieval
Retrieval method using learned dense vector embeddings to find semantically similar documents via approximate nearest neighbour search.
DPR (Dense Passage Retrieval, Facebook 2020) uses a dual-encoder BERT model to embed questions and passages into dense vectors - used in open-domain QA systems to retrieve from 21M Wikipedia passages.
Cohere's embed-multilingual-v3 model powers dense retrieval in 100+ languages across enterprise knowledge bases - a global law firm uses it to retrieve relevant precedents across English, French, and German case law.
NVIDIA cuVS library provides GPU-accelerated approximate nearest neighbour search for dense retrieval - used by vector databases (Milvus, Qdrant) to serve billion-scale embedding search at <10ms latency.