Back to glossary

Retrieval

Bilgi getirme

D3

Retrieval is the process of selecting information items from a collection for use in response to a query or task.

Review status: 2027-08-28

Technical explanation

A retriever can select a small candidate set of passages before another component reads, ranks, or uses those passages to produce an answer.

Conceptual boundaries

Retrieval is not generation and does not require vector embeddings, because systems can use sparse, dense, or other retrieval methods.

Provider-neutral example

Given an employee question, a policy search service can retrieve five candidate sections before a reviewer or model prepares a response.

Limitations

The cited evidence is limited to open-domain question answering and does not show that one sparse or dense retrieval method is universally best.

Related concepts

Atomic claims and evidence

  1. 1.1Karpukhin et al., Dense Passage Retrieval for Open-Domain Question Answering
    Source
    Karpukhin et al., Dense Passage Retrieval for Open-Domain Question Answering
    Source role
    Authoritative source
    Exact locator
    Section 1: two-stage framework with context retriever and machine reader
    Supported claim
    In open-domain question answering, a context retriever selects a small subset of passages before a reader examines the retrieved contexts.
    Last verification
    Review due
    Scope limitation
    This describes a two-stage question-answering framework and does not define every retrieval application.
  2. 2.1Karpukhin et al., Dense Passage Retrieval for Open-Domain Question Answering
    Source
    Karpukhin et al., Dense Passage Retrieval for Open-Domain Question Answering
    Source role
    Authoritative source
    Exact locator
    Abstract; Section 1: sparse TF-IDF/BM25 and dense representations
    Supported claim
    Retrieval can use sparse keyword methods such as TF-IDF or BM25 as well as dense representations.
    Last verification
    Review due
    Scope limitation
    This names methods discussed in the cited research and does not rank one method as universally best.