🤖 AI EngineeringID:
corrective-rag-cragCorrective RAG (CRAG)
Corrective RAG (CRAG) 纠偏检索
🎯Core Definition
Corrective Retrieval-Augmented Generation (CRAG) is a robust RAG architecture that deploys a lightweight Retrieval Evaluator to score the confidence of internal database retrieval; based on calibrated confidence, it triggers one of three branches: 1) Correct (high confidence: runs knowledge refinement filtering before generation); 2) Incorrect (low confidence: completely discards internal passages and falls back to external Web Search APIs); 3) Ambiguous (medium confidence: fuses refined internal passages with external web search results).
💡Use Cases
Production enterprise assistants facing knowledge gaps or stale internal databases, requiring dynamic fallback to live internet search.
⚡Key Problems Solved
Naive RAG forces irrelevant internal noise into LLM context when queries fall outside database coverage, triggering hallucinated falsehoods; CRAG provides automated failure detection and graceful dynamic fallback to web search.
🎯5 High-Frequency Exam Points
1
Describe the confidence classification logic and threshold rules in CRAG's Retrieval Evaluator?
2
How does CRAG's Knowledge Refinement algorithm decompose, filter, and recompose fine-grained passage contents?
3
When falling back to Web Search (Tavily/Bing), how to clean, filter, and rerank live web search snippets on the fly?
4
Compare CRAG's modular 3-branch pipeline vs Self-RAG's internal token generation in production maintainability?
5
In air-gapped corporate networks without internet access, how should CRAG gracefully degrade when internal retrieval fails?
📖 In-depth Guide:📄 naive-and-advanced-rag →
Updated 2026-08-14🎯
🚀 Start Card Practice ➔Test Your Knowledge: Practice Questions for "Corrective RAG (CRAG)"
Single choice pitfall questions with instant feedback and mistake tracking.