1Cademy - Single-Vector Dense Passage Retrieval

Learn Before

Research

Concept

Single-Vector Dense Passage Retrieval

Single-vector dense passage retrieval is a neural information-retrieval paradigm in which a query and each candidate passage are independently mapped to a single fixed-dimensional dense embedding by neural encoders, and relevance is scored as a similarity (typically the dot product or cosine similarity) between those two pooled vectors. Because each passage is represented by exactly one vector, the corpus can be pre-encoded once and indexed with a maximum-inner-product search (MIPS) structure such as FAISS, so retrieval at query time reduces to a nearest-neighbor lookup against the query embedding. This contrasts with multi-vector / late-interaction retrievers, which keep one contextualized embedding per token and defer fine-grained matching to query time at the cost of much larger indexes. Single-vector dense retrievers are the standard 'flat dense retrieval' baseline in modern retrieval-augmented generation pipelines.