Corrective RAG (CRAG)
Corrective RAG adds self-evaluation and correction mechanisms to traditional RAG, using LangGraph to implement a sophisticated workflow that validates retrievals and uses fallback strategies when needed.Overview
CRAG evaluates the quality of retrieved documents and takes corrective actions:- Relevant documents: Use directly for generation
- Ambiguous documents: Apply query transformation and re-retrieve
- Irrelevant documents: Fall back to web search
CRAG significantly improves answer accuracy by validating retrieval quality before generation.
Architecture
Implementation
See the Advanced RAG Techniques page for complete CRAG implementation with LangGraph.Document Grading
Evaluate retrieval relevance with LLM-based grading
Query Transformation
Rewrite queries for better retrieval on ambiguous results
Web Search Fallback
Use Tavily AI for web search when local docs insufficient
LangGraph Workflow
Orchestrate the complete CRAG workflow with state management
Related Examples
Advanced RAG Techniques
Complete CRAG implementation with code examples and LangGraph workflow