SmartFAQs.ai
Back to Learn
Intermediate

Corrective RAG (CRAG)

Corrective RAG (CRAG) is an architectural pattern that introduces a self-correction mechanism by using a retrieval evaluator to score the relevance of retrieved documents; it triggers supplementary web searches if the local vector store results are deemed insufficient. This increases reliability and minimizes hallucinations at the cost of increased latency and higher API expenditure due to additional LLM evaluation steps and external search calls.

Definition

Corrective RAG (CRAG) is an architectural pattern that introduces a self-correction mechanism by using a retrieval evaluator to score the relevance of retrieved documents; it triggers supplementary web searches if the local vector store results are deemed insufficient. This increases reliability and minimizes hallucinations at the cost of increased latency and higher API expenditure due to additional LLM evaluation steps and external search calls.

Disambiguation

Distinguished from standard RAG by its ability to 'reject' local retrieval and pivot to external data sources.

Visual Metaphor

"A library researcher who, upon finding only outdated or irrelevant books on a shelf, leaves the building to consult a global news archive."

Key Tools
LangGraphTavilyLlamaIndexSerperLangChain
Related Connections

Conceptual Overview

Corrective RAG (CRAG) is an architectural pattern that introduces a self-correction mechanism by using a retrieval evaluator to score the relevance of retrieved documents; it triggers supplementary web searches if the local vector store results are deemed insufficient. This increases reliability and minimizes hallucinations at the cost of increased latency and higher API expenditure due to additional LLM evaluation steps and external search calls.

Disambiguation

Distinguished from standard RAG by its ability to 'reject' local retrieval and pivot to external data sources.

Visual Analog

A library researcher who, upon finding only outdated or irrelevant books on a shelf, leaves the building to consult a global news archive.

Related Articles