Corrective RAG
Authoritative Definition
A robust Retrieval-Augmented Generation strategy that dynamically evaluates the relevance of retrieved documents and triggers fallback actions, such as web searches or query rewrites, when confidence is low.
Overview & Technical Description
Corrective Retrieval-Augmented Generation (CRAG) introduces an explicit evaluation phase between the retrieval and generation steps in standard RAG architectures. By employing a lightweight evaluator model or a specific prompt, CRAG scores retrieved chunks for relevance against the users query. If the retrieved internal documents score highly, generation proceeds normally. If the confidence score falls below a threshold, CRAG initiates corrective actions. These actions may include rewriting the original query for better dense retrieval, dropping irrelevant chunks, or routing the query to external search engines (like Google or Tavily) to fetch the missing information. This self-correcting loop significantly reduces hallucinations caused by poor retrieval quality.
Editorial Notes
Prevents low-quality retrieval from degrading generation. CRAG is highly effective in production systems where internal knowledge bases may be incomplete, seamlessly supplementing with web data when necessary.