01

What changed

The original research combined a pretrained text generator with a searchable collection of Wikipedia passages. A neural retriever selected passages; the generator used them alongside the question to produce its output. The index represented passages as numerical vectors for retrieval. In a separate experiment within that paper, the researchers reported changing the system's answers about world leaders by swapping Wikipedia indexes from different years without retraining the model. That is a research result, not a promise that every deployment updates itself. Full paper

02

Why it matters

RAG separates some of the information a system can consult from the model that writes its answers. The external collection can be changed, and the retrieved text can be inspected. In the original research, that collection was a Wikipedia index, not an automatic live web search. Editorial analysis: assess whether an assistant finds appropriate evidence and uses it faithfully. Full paper

03

Who it affects

This explainer is for working readers evaluating document-based assistants or encountering RAG in an AI tool description. It is especially relevant when answers need to reflect a defined set of documents rather than general training knowledge.

04

What you can do with this

Editorial next step: choose a small document set you know well and write a few questions with known answers, plus one whose answer is absent. If the assistant exposes retrieved passages, compare those passages with the original documents, then check whether the generated answer is supported by them. Check source dates and ask the provider whether supporting passages are visible. For the missing-answer question, observe whether the assistant admits it lacks evidence instead of guessing. These are suggested assessment steps, not tests performed here or a validated benchmark; the original paper establishes the separate retrieval and generation stages. Full paper

05

What we don’t know

Retrieval alone does not guarantee relevant, accurate or current evidence, and generation does not guarantee faithful use of it. The RAG researchers reported preliminary experiments on some tasks, including story generation, where retrieval returned the same documents regardless of the input and the generator learned to ignore them. Full paper A separate study first submitted in July 2023 found that tested models often used relevant information less effectively when it appeared in the middle of long inputs. Study record These are findings about particular research systems and experiments, not measurements of every model available in November 2026 or proof that RAG eliminates invented answers.

Correction · Nov 19, 2026

Updated the source-status description and stated the editorial assessment directly. Historical research evidence, citation destinations and limitations are unchanged.

SHOWING OUR WORK

Go straight to the source.

Facts are grounded in the references below. Practical implications are editorial analysis unless explicitly attributed to a source.

01
Lewis et al. — Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks, original arXiv recordarxiv.org · May 22, 2020
02
Lewis et al. — Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks, full paper v4arxiv.org · Apr 12, 2021
03
Liu et al. — Lost in the Middle: How Language Models Use Long Contexts, original arXiv recordarxiv.org · Jul 6, 2023
See claim-level references

The original RAG paper was first submitted to arXiv on May 22, 2020, accepted at NeurIPS 2020, and last revised as v4 on April 12, 2021. Source

The original RAG models combined a pretrained text generator with a dense vector index of Wikipedia accessed by a neural retriever. Source

The original RAG generator uses retrieved text together with its input to generate output; it is not limited to returning an extracted search passage. Source

In the original research, the external collection was an indexed Wikipedia snapshot, not automatic live web retrieval. Source

The authors report changing answers about world leaders by replacing Wikipedia indexes from different years without retraining; this is a scoped research finding, not automatic freshness for all deployments. Source

The paper describes its external text memory as human-readable and human-writable, supporting inspection and updates to the document index. Source

Appendix H reports retrieval collapse in preliminary experiments on some tasks, including story generation: the same documents were retrieved regardless of input, and the generator learned to ignore them. Source

Lost in the Middle was first submitted on July 6, 2023, and reports that tested models often performed worse when relevant information appeared in the middle of long input contexts. Source