Corrective RAG (CRAG) — AI Glossary

Reference details

Related terms

• Self-RAG
• Advanced RAG
• Retrieval Pipeline

How it works

1. Retrieval runs first, and a lightweight evaluator grades the returned documents rather than passing them straight to generation.
2. The grade is categorical — broadly correct, ambiguous, or incorrect — because a continuous score still needs a decision boundary somewhere.
3. On a correct grade the documents are refined: decomposed into relevant strips and recomposed, discarding the parts that matched but do not bear on the question.
4. On an incorrect grade the pipeline goes elsewhere, typically to web search, rather than generating from poor context.
5. On ambiguous it does both and combines, which is the hedge against a grader that is itself imperfect.

Example: The grader's accuracy bounds the benefit

Suppose 20% of retrievals are poor, and generating from poor context yields a wrong answer. A grader that catches 90% of those converts 18 of every 100 requests from wrong to corrected, at the cost of one extra evaluation call on all 100 plus a fallback on 18. A grader catching only 50% converts 10 and still charges the evaluation on all 100. The method's value scales with grader recall, which is why a cheap grader that misses half the bad retrievals barely repays its own overhead.

Commonly confused with

Grades retrieved documents for relevance and, when they fail, goes somewhere else — typically web search — instead of generating from poor context. The trigger is what separates it: Self-RAG reflects on its own generation, agentic RAG decides retrieval strategy up front, and this reacts to a measured retrieval failure. It presumes a fallback source exists, which is a real deployment constraint.

When to use it

Reach for it when:

• Retrieval quality varies and generating from poor context produces confident wrong answers.
• A fallback source exists that can cover what the corpus misses.
• The extra evaluation call per request is affordable.

Reach for something else when:

• Retrieval is reliably good, where the grader is pure overhead.
• No fallback source exists, so a poor grade has nowhere to go.
• The grader has not been measured, leaving its recall — and therefore the benefit — unknown.

Primary source

• Yan et al., Corrective Retrieval Augmented Generation (arXiv)

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Corrective RAG (CRAG) — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path