What is Corrective RAG (CRAG)?
Corrective RAG (CRAG) is an advanced retrieval-augmented generation pattern that incorporates self-correction mechanisms to evaluate and improve the quality of retrieved information before using it for generation. Unlike standard RAG that assumes retrieved documents are relevant, CRAG includes an evaluation step where the system assesses whether the retrieved content actually addresses the query and takes corrective actions if the retrieval quality is insufficient.
The corrective mechanism typically involves several strategies: if retrieved documents are assessed as highly relevant, they are used directly; if they are somewhat relevant, they may be further processed or filtered to extract the most pertinent information; if they are deemed irrelevant, the system may reformulate the query, expand the search to additional sources, or fall back to alternative retrieval strategies. Some implementations use web search as a fallback when internal knowledge bases don't yield good results.
CRAG represents an important evolution in RAG systems, addressing a common failure mode where irrelevant or low-quality retrievals lead to poor generations. By adding self-evaluation and correction loops, CRAG systems become more robust and reliable, particularly for queries that are ambiguous, complex, or outside the coverage of the primary knowledge base. This approach aligns with broader trends toward more agentic and self-improving AI systems.