Mehmet EmrahKonyaTR
← Back to glossary

Retrieval-augmented generation (RAG)

Bilgi getirim destekli üretim (RAG)

D3

Retrieval-augmented generation is a generation approach that combines a model’s parametric knowledge with retrieved external information.

Review status: 2026-11-26

Technical explanation

A RAG application can retrieve documents or other resources at runtime and add selected material to the model context before it generates a response.

Conceptual boundaries

RAG is not retrieval alone, does not necessarily use vector search, and does not guarantee grounding, provenance, or factual correctness.

Provider-neutral example

Before drafting an answer about a policy, an application can retrieve relevant policy sections and supply them with the question to a language model.

Limitations

Retrieved material can be untrusted or wrong, and the original RAG study reports comparative results rather than a guarantee that every RAG answer is factual.

Related concepts

Atomic claims and evidence

  1. 1.1Lewis et al., Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
    Source
    Lewis et al., Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
    Source role
    Authoritative source
    Exact locator
    Abstract; Section 1: parametric/non-parametric memory and dense Wikipedia index
    Supported claim
    The original RAG paper defines RAG models as combining pre-trained parametric and non-parametric memory for language generation, with a neural retriever accessing an external index in its proposed formulation.
    Last verification
    Review due
    Scope limitation
    The neural-retriever and dense-index details describe the paper’s formulation and do not prescribe every later RAG implementation.
  2. 2.1NIST AI 100-2e2025, Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations
    Source
    NIST AI 100-2e2025, Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations
    Source role
    Authoritative source
    Exact locator
    Section 3.1.1 printed p. 39; Figure 6 printed p. 39
    Supported claim
    NIST describes RAG application context as query-dependent runtime material from external sources and recommends reviewing source and citation claims in GAI outputs.
    Last verification
    Review due
    Scope limitation
    This is risk-management guidance and does not assert that every RAG system uses the same pipeline or produces incorrect output.
    2.2NIST AI 600-1, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile — D3 evidence slice
    Source
    NIST AI 600-1, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile — D3 evidence slice
    Source role
    Authoritative source
    Exact locator
    MEASURE 2.5, MS-2.5-003 printed p. 32: review and verify sources and citations in GAI outputs
    Supported claim
    NIST describes RAG application context as query-dependent runtime material from external sources and recommends reviewing source and citation claims in GAI outputs.
    Last verification
    Review due
    Scope limitation
    This is risk-management guidance and does not assert that every RAG system uses the same pipeline or produces incorrect output.