What are Corrective RAG, Self-RAG, and Adaptive RAG?

30-second answerSay your answer out loud first, then reveal.

They add decision-making and self-checking around retrieval. Adaptive RAG routes each query to no retrieval, single-shot retrieval, or iterative retrieval based on complexity. Corrective RAG (CRAG) grades retrieved documents and, if they're poor, rewrites the query or falls back to another source such as web search. Self-RAG trains or prompts the model to decide when to retrieve and to critique its own output for relevance and support. They're the building blocks of agentic RAG.

An adaptive router sends a query to a direct answer, single retrieval or an iterative loop; retrieved docs are graded, poor ones trigger a query rewrite or another source, and the generated answer is checked for grounding before it is returned.

How the three compare

Approach	Key idea	Cost
Adaptive RAG	Classifier picks the pipeline per query	Saves cost on simple queries
Corrective RAG (Yan et al. 2024)	Retrieval evaluator; corrective actions (refine, web search)	Extra grading calls
Self-RAG (Asai et al. 2023)	Model emits reflection tokens: retrieve? relevant? supported? useful?	Originally needs a fine-tuned model; prompt-based versions exist

Practical implementation: a state graph (e.g. LangGraph) with nodes for route → retrieve → grade → generate → check, and edges for retries. Cap the loops.

Trade-offs: better robustness at the cost of latency (2–5x more LLM calls) and complexity. Use them for high-value or complex queries; keep simple queries on the fast path.

PreviousHow do you build multimodal RAG over documents with images, charts, and scanned pages?

Every expert started right here.

What are Corrective RAG, Self-RAG, and Adaptive RAG?

How the three compare

Related