Definition
Corrective Retrieval-Augmented Generation (CRAG) is an architectural strategy that utilizes a dedicated retrieval evaluator to assess the relevance of retrieved documents, enabling the pipeline to trigger corrective actions—such as web searches or document filtering—when local knowledge retrieval is deemed insufficient or inaccurate.
Not just basic RAG; it's a self-correcting loop that dynamically switches between local knowledge and external search based on confidence scores.
"A gatekeeper at a factory intake valve who tests raw materials for purity and diverts them to a purification tank if they fail the quality check."
- RAG(Prerequisite)
- Self-RAG(Advanced Sibling)
- Retrieval Evaluator(Core Component)
- Knowledge Refinement(Process Step)
Conceptual Overview
Corrective Retrieval-Augmented Generation (CRAG) is an architectural strategy that utilizes a dedicated retrieval evaluator to assess the relevance of retrieved documents, enabling the pipeline to trigger corrective actions—such as web searches or document filtering—when local knowledge retrieval is deemed insufficient or inaccurate.
Disambiguation
Not just basic RAG; it's a self-correcting loop that dynamically switches between local knowledge and external search based on confidence scores.
Visual Analog
A gatekeeper at a factory intake valve who tests raw materials for purity and diverts them to a purification tank if they fail the quality check.