MMOOZZGG.com

Glossary

What is corrective RAG?

In short

Corrective RAG (CRAG) is a document search scheme in which a separate model first rates what was found and fixes poor results: it drops what does not fit or searches other sources.

Search and RAG · CRAG

MMOOZZGG.com editorsChecked: 1 min read

For a child

You come to the library for a book about pirates, and the librarian gives you a book about pirate toys by mistake. A careful student first flips through it and sees it is the wrong one. Then the student asks for another book or looks in an encyclopedia. Corrective RAG teaches a computer to check what it was given in the same way before it answers a question.

In plain words

The usual scheme trusts the search: whatever it finds, the model reads. If it finds the wrong thing, the answer comes out confident and wrong. Corrective RAG puts a checker between the search and the answer. A small model looks at each passage found and decides whether it fits the question. If it fits, only the useful sentences are kept. If everything misses, the system searches another source. If it is unclear, it takes both. For example, asked about a new rate plan that is not in the database yet, the assistant will not retell the old price list.

For business

The scheme helps when the knowledge base is incomplete or ages faster than it is updated: catalogs, rate plans, changes in the law. The checking model is small, so it adds little delay and cost, and the main gain is that poor results do not turn into a confident answer. The risk is in the web search step: outside sites can be wrong, and a query sent outside reveals what staff are asking about. For internal tasks, this step is often replaced by a search in a second database or an honest answer: “the documents do not cover this”. The checker’s thresholds are tuned on your questions; this is a normal part of setting up an AI assistant for documents.

For a specialist

Corrective Retrieval Augmented Generation was proposed by Yan, Gu, Zhu and Ling (2024) as an add-on that works with different RAG pipelines. A lightweight retrieval evaluator based on T5-large, about 0.77 billion parameters, gives each question-document pair a score from -1 to 1. Two thresholds pick the action. Correct: at least one document is above the upper threshold, and the internal knowledge is refined. Incorrect: all are below the lower threshold, so the documents are dropped and replaced with web search results, for which the input is rewritten into a keyword query. Ambiguous: both sources are used. Refinement uses a decompose-then-recompose algorithm: a document is split into knowledge strips of a few sentences, the evaluator removes irrelevant ones, and the rest are joined in order. The experiments used PopQA, Biography, PubHealth and Arc-Challenge, with thresholds chosen empirically for each set. The authors call the version built on top of Self-RAG Self-CRAG.

Example

An employee asks about a rate plan launched yesterday, but the knowledge base will update only in a week. The checking model sees that the passages found describe old rate plans, and instead of a confident wrong answer, the system searches the company website.

Related terms

Questions and answers

How is corrective RAG different from Self-RAG? In Self-RAG, the checks are built into the generating model itself. In corrective RAG, a separate small model rates the results, so you can add the scheme to a pipeline that already works.
Do you have to search the internet? No. In the original paper, poor results trigger a web search, but companies often replace it with a second internal database or the answer “the documents do not cover this”.

That way, information about staff questions does not leave the company.

Sources

Where companies use it

Not sure whether your company needs this?

AI solutions map
Contact us