# Agentic RAG > VI: https://5ac.vn/wiki/agentic-rag/ · EN: https://5ac.vn/en/wiki/agentic-rag/ · lang: en · spec_as_of 2026-10-18 (spec no formal spec) · cluster C17 A RAG architecture that embeds autonomous agents in the retrieval pipeline: the agent decides what to search, in which sources, and how many passes, instead of one fixed vector lookup. --- ## Definition Agentic RAG is a retrieval-augmented generation architecture that embeds autonomous agents in the retrieval pipeline: instead of one static similarity lookup over fixed chunks, the agent decides what to search, in which sources, and how many passes to make, steered by four design patterns — reflection, planning, tool use and multi-agent collaboration. ## Mechanism Lewis et al. (2020) set the classical form: a pretrained seq2seq model's parametric memory fused with non-parametric memory (a dense vector index of Wikipedia reached through a neural retriever), tested in two conditioning styles, one passage set for the whole answer or a different passage per token. The resulting shape (ask once, fetch the most similar chunks, generate) is exactly the ceiling that the Agentic RAG survey names: "traditional RAG systems are constrained by static workflows and lack the adaptability required for multi-step reasoning and complex task management." Agentic RAG breaks that shape by putting agents inside the pipeline. Singh et al. describe agents exploiting four design patterns (reflection, planning, tool use, multi-agent collaboration) to dynamically manage retrieval strategy, iteratively refine contextual understanding, and reshape workflows from sequential steps to adaptive collaboration. The same survey taxonomizes architectures along four axes: agent cardinality, control structure, autonomy, and knowledge representation — enough coordinates to place any concrete design on the map. | Pattern | Job inside the retrieval pipeline | | --- | --- | | Reflection | Inspects intermediate results, spots missing or noisy context, redirects the search | | Planning | Decomposes the question into retrieval steps, orders them, sets the stop condition | | Tool use | Calls real sources — web search, databases, filesystems — not one vector store alone | | Multi-agent collaboration | Splits branches of retrieval across parallel agents, then merges findings | On the production side, Anthropic draws the same boundary when contrasting its Research system with RAG: the traditional approach fetches the chunks most similar to the input once and generates from them (static retrieval) while its architecture runs multi-step search: find, read what came back, adapt to new findings, analyze results before answering. Applications the survey walks through: healthcare, finance, education, enterprise document processing. Open research challenges it lists: evaluation, coordination, memory management, efficiency and governance — token cost of retrieval loops being the number every agentic design has to answer before shipping. Flow: Question → Agent plans retrieval → Iterative search across sources → Evaluate and refill context → Answer with sources ## Spec anchor No standards body publishes Agentic RAG; it is an academic label, so this anchor pins source-document revisions instead of a spec version. Revision numbers below were read from the arXiv API on 2026-10-18: the survey stands at v4 (updated 2026-04-01), the original RAG paper at v4 (updated 2021-04-12). When a third-party article cites "agentic RAG," check which survey revision it mirrors — taxonomy and application lists shift between revisions. Sources fetched for this entry: - — Agentic RAG survey abstract page (v4) — accessed 2026-10-18 - — original RAG paper abstract page (v4) — accessed 2026-10-18 - — arXiv API: verified IDs, authors, abstracts, revision dates — accessed 2026-10-18 - — static retrieval vs multi-step search contrast — accessed 2026-10-18 ## Academic anchors Both IDs are pinned in the frontmatter and were verified (ID + authors + abstract) through the arXiv API on 2026-10-18. *Agentic Retrieval-Augmented Generation: A Survey on Agentic RAG* (arXiv:2501.09136, Aditi Singh, Abul Ehtesham, Saket Kumar, Tala Talaei Khoei, Athanasios V. Vasilakos; submitted 2025-01-15) is the direct source of this entry's definition, four-pattern set and taxonomy axes. *Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks* (arXiv:2005.11401, Patrick Lewis, Ethan Perez, Aleksandra Piktus et al.; submitted 2020-05-22) is the baseline to compare against — its own abstract describes the static machinery the agentic line was born to unlock. ## FAQ **What separates agentic RAG from traditional RAG?** Traditional RAG retrieves statically: one pass over the chunks most similar to the question, then generate. Agentic RAG turns retrieval into a deliberate loop — the agent evaluates intermediate evidence, changes strategy, searches again or switches sources until the ground is sufficient. **When is it worth upgrading from one-shot RAG?** When questions need multi-step reasoning, when the data lives in heterogeneous sources rather than one pre-vectorized corpus, and when better answers justify the extra token spend of several retrieval passes. **Where do the four patterns in the definition come from?** From the survey itself: reflection, planning, tool use and multi-agent collaboration are the agentic design patterns it credits with making retrieval dynamic — each tied to a concrete decision (correct the search direction, decompose the query, pick sources, split parallel branches). **Which open problems does the survey flag?** Five groups: evaluation (no shared comparison standard yet), inter-agent coordination, memory management across loops, cost efficiency, and governance. That is also why this entry carries no benchmark numbers — no runner plus benchmark version has been pinned yet. ## Related - [Open source AI infrastructure for enterprises: Hermes, Gbrain and VPS self-host](/en/blog/open-source-ai-infrastructure-self-hosted/) - [What is agent memory?](/en/wiki/agent-memory/) - [What is a sub-agent?](/en/wiki/sub-agent/) - [What is a Multi-Agent OS?](/en/blog/what-is-multi-agent-os/)