Applications and Future of Dense Retrieval in Industry

Yubin Kim

Abstract: Large-scale search engines are often designed as tiered systems with at least two layers. The L1 candidate retrieval layer efficiently generates a subset of potentially relevant documents (typically ~1000 documents) from a corpus many orders of magnitude larger in size. L1 systems emphasize efficiency and are designed to maximize recall. The L2 re-ranking layer uses a more computationally expensive, but more accurate model (e.g. learning-to-rank or neural model) to re-rank the candidates generated by L1 in order to maximize precision of the final result list. Traditionally, candidate retrieval was performed with an inverted index data structure, with exact lexical matching. Candidates are ordered by a dot-product-like scoring function f(q,d) where q and d are sparse vectors containing token weights, typically derived from the token's frequency in the document/query and corpus. The inverted index enables sub-linear ranking of the documents. Due to the sparse vector representation of the documents and queries, lexical match retrieval systems have also been called sparse retrieval. To contrast, dense retrieval represents queries and documents by embedding the text into lower dimensional dense vectors. Candidate documents are scored based on the distance between the query and document embedding vectors. Practically, the similarity computations are made efficiently with approximate k-nearest neighbours (ANN) systems. In this panel, we bring together experts in dense retrieval across multiple industry applications, including web search, enterprise and personal search, e-commerce, and out-of-domain retrieval.