Corrective RAGdefinition and how it works in 2026
- Corrective RAG
- A RAG variant that grades retrieved documents and triggers fallback retrieval (web search, alternative sources) when the initial retrieval scores low on relevance.
Corrective RAG (CRAG) sits between the retrieval step and the generation step. It uses a lightweight retrieval evaluator to grade docs as "correct," "ambiguous," or "incorrect." Correct docs proceed normally. Incorrect docs trigger a fallback — usually web search or a different corpus. Ambiguous docs may trigger both for triangulation.
The technique addresses standard RAG's biggest failure mode: confidently generating from irrelevant retrieved docs. CRAG's evaluator catches the mismatch and reroutes, dramatically reducing hallucinations on queries where your primary corpus doesn't have the answer.
Most production agentic-RAG platforms in 2026 implement some form of corrective retrieval — usually via reranker confidence scores triggering a fallback path rather than the full CRAG architecture from the original paper.
Frequently asked
What's the difference between Corrective RAG and a reranker?+
A reranker scores and reorders retrieved docs. CRAG adds a binary decision (correct/incorrect) that triggers fallback retrieval. CRAG often uses a reranker internally as the evaluator.
When does Corrective RAG matter most?+
When your primary corpus has gaps (e.g. recent events, niche topics) but the queries don't respect them. CRAG's fallback to web search rescues the queries that would otherwise return confident nonsense.