Agentic RAG
Retrieval-augmented generation in which an agent plans, retrieves, evaluates and re-retrieves iteratively, instead of fetching context once.
- term
- Agentic RAG
- category
- knowledge-memory
- short_def
- Retrieval-augmented generation in which an agent plans, retrieves, evaluates and re-retrieves iteratively, instead of fetching context once.
- long_def
- Where naive RAG runs a single similarity search and hands the results to the model, agentic RAG turns retrieval into a control loop: the agent decides when and what to retrieve, judges whether the results are sufficient, retries or switches tools (web search, SQL, APIs) and validates before answering. It trades latency and cost for reliability on complex, multi-step questions.
- see_also
ragembeddingsvector-database- etymology_origin
- — verify-against-primary-at-build ↗ https://arxiv.org/abs/2501.09136 — 'agentic RAG' is a 2024-2025 practitioner/research term building on RAG (Lewis et al. 2020); see the 2025 survey arXiv 2501.09136, with no single coiner for the exact term
- related_to
ragembeddingsvector-databaseagentic-loopagent-skills- contrast_with
- Unlike naive (one-shot) RAG, which retrieves context a single time before generating, agentic RAG defers retrieval decisions to an agent that can retrieve, evaluate, re-retrieve and validate in a loop — a control loop, not a fixed pipeline.
- example
- A 2025 survey on agentic RAG (arXiv 2501.09136) frames it as turning retrieval from a static pipeline into an iterative, agent-driven control loop.
- source
- https://arxiv.org/abs/2501.09136
- status
- emerging
- why_it_matters
- Agentic RAG is how modern agents ground complex answers; sites that expose clean, chunkable, well-described content are easier for an iterating retrieval agent to use and cite.
- sameAs
- —
- bridge_entity
- models
- last_verified
- 2026-06-15
- md_twin
- /glossary/agentic-rag.md