Corrective RAG (CRAG): Self-Evaluation and Correction

Trusting retrieval less, on purpose

Every architecture so far hands retrieved chunks straight to the language model and trusts them. Corrective RAG (CRAG) inserts a deliberate check between retrieval and generation: a lightweight evaluator model grades the retrieved chunks, typically as correct, ambiguous, or incorrect, before generation is allowed to proceed.

What happens on each grade

On a correct grade, CRAG behaves like any other RAG system: the chunks go straight to generation. On an incorrect grade, the system does not generate from that evidence at all; instead it falls back to an alternative source, most commonly a live web search, to fetch better evidence before trying again. On an ambiguous grade, CRAG does both: it keeps the useful parts of the original retrieval and supplements them with an external search, rather than discarding potentially-good evidence just because it was not a clean match.

Why this matters for reliability

CRAG's core insight is that retrieval failure and generation failure are separable problems, and treating them separately makes both easier to fix. A system that only ever gets better at generation (better prompts, better models) has no defense against being handed genuinely bad evidence, it will confidently write a fluent answer from irrelevant chunks. Grading the evidence first, and having a defined fallback for a bad grade, catches the retrieval failure before it becomes a generation failure a user actually sees.