Corrective RAG (CRAG): Self-Evaluating Retrieval
~12 min read
Corrective RAG adds a validation step after retrieval: check retrieved results against trusted sources before trusting them, and fall back to a web search when the vector store's results aren't good enough.
Corrective RAG (CRAG): Self-Evaluating Retrieval is a Pro topic
Sign in, then upgrade to Pro or Power to unlock this topic and the full AI Engineering curriculum.
Key points
- •CRAG adds a relevance-evaluation step between retrieval and generation — retrieved content isn't automatically trusted
- •The evaluator scores retrieved chunks for genuine relevance/trustworthiness, not just cosine similarity
- •Sufficiently relevant content proceeds like naive RAG; insufficient content triggers a fallback to an external source like web search
- •Especially valuable for fast-changing information, where a vector store's static content can be stale even when it's the closest semantic match
- •Trade-off: more robust retrieval at the cost of an extra evaluation step (and possibly a web search call) in the pipeline