Corrective Retrieval Augmented Generation

by Beihang University / University of Illinois Urbana-Champaign / Microsoft · free · Last verified 2026-03-17

Introduces Corrective RAG (CRAG), which incorporates a lightweight retrieval evaluator to assess the relevance of retrieved documents and triggers corrective actions including web search when retrieved documents are deemed irrelevant or ambiguous. CRAG improves robustness of RAG systems against imperfect retrieval across diverse datasets.