EXPLAINER

Agentic RAG (Adaptive and Self-Correcting Retrieval)

techniquefreshness: evolvingsources not yet verified

Agentic RAG lets a model decide when and how to retrieve — rewriting queries, choosing sources, checking relevance and retrying — instead of a single fixed retrieve-then-generate pass.

Short answer: The retrieval step becomes a decision loop: the model plans queries, calls search tools, grades what came back, and retries or stops. It handles complex questions better but costs more and adds failure modes; use it only where plain RAG falls short.

What is it?

Basic RAG always searches once. Agentic RAG lets the AI act like a researcher: reword the question, try a different source, notice weak results and try again before answering.

Also known as: agentic RAG, self-RAG, corrective RAG, CRAG, adaptive RAG, iterative retrieval, multi-hop RAG, query decomposition.

Technical definition

Patterns: query rewriting/decomposition, routing among indexes or tools, relevance grading with retry (Corrective RAG, Yan et al., 2024), self-reflective retrieval where the model decides when to retrieve (Self-RAG, Asai et al., 2023), and iterative multi-hop retrieval. It combines What is RAG (Retrieval-Augmented Generation)? with The ReAct Pattern (Reason + Act). Extra LLM calls add latency and cost, and a mis-judging grader can discard good evidence.

Entity type: technique · Freshness class: evolving.

When should I use it?

Use when single-pass retrieval fails on multi-hop or ambiguous questions.

Implementation patterns

  • Rewrite → retrieve → grade → (retry | answer).
  • Router selects among vector, keyword and web search.
  • Decompose into sub-questions; answer; synthesise.
  • Hard cap on retrieval iterations.

Technologies and frameworks

Commonly associated: LangGraph, LlamaIndex agents, search tools, rerankers. Names are examples of what exists, not recommendations, and projects change quickly.

Practical guide

  1. Measure plain-RAG failures first.
  2. Add the smallest loop that fixes them.
  3. Limit iterations.
  4. Evaluate end-to-end and retrieval separately.

Common failure modes

  • Loops that never stop.
  • Grader errors discarding useful passages.
  • Latency and cost multiplication.
  • Injected instructions in retrieved text (What is Prompt Injection?).

Troubleshooting

  • Slow answers — Cap iterations, parallelise sub-queries, use a small grader model.
  • Answers still ungrounded — Require citations and verify them against retrieved text.

Contrasted with

Questions this page answers

  • What is agentic RAG?
  • What is self-RAG?
  • When should RAG retrieve more than once?
  • How do I make RAG handle multi-hop questions?

Typical goals: handle complex multi-part questions; add self-checking to RAG.

Sources and verification

Status: not independently verified. This entry was authored on 2026-10-06 from general technical knowledge, without live checking of the sources below. Treat every source as a lead to confirm (identifier, title, authorship, current version) before publication or citation.

Freshness class: evolving — Established techniques whose best practice shifts as research and tooling move. Re-review every 6 months.

Canonical sources (unverified leads):

  • Asai et al., Self-RAG (2023) — arXiv:2310.11511
  • Yan et al., Corrective Retrieval Augmented Generation (2024) — arXiv:2401.15884