botminds.aiRequest a demo
Guide · Agentic RAG

Agentic RAG vs RAG: a fixed pipeline against a loop that decides.

Classic RAG retrieves once for every question. Agentic RAG lets an agent choose whether to search, where, which way — and whether the result is good enough.

Retrieval augmented generation gives a language model passages from your documents so it answers from them instead of from memory. This guide shows how the classic pipeline works, the loop an agent adds on top, the five years of research between them, a worked trace of one question, and how to tell when the extra steps are worth their cost.

Updated · 7 sources

Read the guideHow Botminds does it
retrieve, then generate

RAG

A fixed pipeline: embed the question, fetch the nearest passages from one index, put them in the prompt, generate.

  • One search per question
  • One index, one search method
  • Whatever comes back is used
  • Cheap, fast, predictable
plan, search, check, repeat

Agentic RAG

An agent runs retrieval as a set of tools: it plans, picks a source and a method, grades what it found and searches again when the passages fall short.

  • Zero, one or many searches
  • Many sources and search methods
  • Results judged before use
  • More calls, better on hard questions
The pipeline, every timeWho decides to searchThe agent, per question
The user’s words, embeddedQueryRewritten, split into sub-questions
One vector indexSourcesIndexes, databases, documents, the web
Passed to the model anywayBad retrievalDetected; search again or say so
One retrieval, one model callCost per answerSeveral tool calls and model calls

What is the difference between agentic RAG and RAG?

In classic RAG, a fixed pipeline retrieves passages once for every question and hands them to the model. In agentic RAG, an AI agent runs retrieval as tools: it decides whether a question needs sources, which source and search method to use, checks whether the passages answer the question, and searches again or says so when they do not. Agentic RAG vs RAG is the difference between a single lookup and a research loop.

The baseline

How classic RAG works: index once, retrieve once, generate.

The 2020 paper that named retrieval augmented generation paired a language model’s built-in (parametric) memory with an external (non-parametric) memory: a dense vector index of Wikipedia, searched by a neural retriever. Production RAG still follows that shape.

  1. Chunk

    Documents are cut into passages of a few hundred words.

    offline
  2. Embed

    Each passage becomes a vector with an embedding model.

    offline
  3. Index

    Vectors go into an index built for nearest-neighbour search.

    offline
  4. Embed the question

    At question time, the same model embeds the user’s words.

    per question
  5. Fetch top-k

    The k nearest passages are returned — relevant or not.

    per question
  6. Generate

    Passages and question go into one prompt; the model answers.

    per question

Where the fixed pipeline breaks

  • Multi-part questionsOne embedding of a two-part question finds passages for one part.
  • Exact identifiersA clause number or error code blurs into similar-looking text.
  • Wrong sourceThe answer lives in a table or a database the index never held.
  • Nothing relevantTop-k always returns k passages, and the model tries to use them.

Naive, advanced, modular. A widely cited 2023 survey sorts RAG systems into three paradigms: naive RAG (the six steps here), advanced RAG (adds query rewriting and re-ranking around retrieval) and modular RAG (swappable modules, including routing and iteration). Agentic RAG hands the control of those modules to an agent.

What the agent adds

How agentic RAG works: the retrieval loop an agent runs.

An agentic RAG system gives the model search tools and lets it reason between calls. The loop below runs until the agent has enough evidence or decides the sources do not cover the question.

One question, inside an agentic RAG agent
  1. PlanDoes this need sources at all? If so, split it into sub-questions.
  2. RoutePick the source — a document library, a records table, the web — for each sub-question.
  3. RetrieveChoose the method: by meaning, by exact words, by filter, or a database query.
  4. GradeDo these passages actually answer the sub-question?
  5. AnswerCompose from graded evidence, cite the passages used, or say what is missing.

Passages fall short: rewrite the query, try another source or method, and go round again.

A worked trace: “Which of our EU suppliers had late deliveries last quarter, and what do their contracts say about penalties?”
plan      → two sub-questions: (1) late EU deliveries, (2) penalty clauses
route (1) → deliveries table, not the document index
query (1) → region = EU AND delivered_after_due AND quarter = Q3
          ← 3 suppliers
route (2) → contracts library, filtered to those 3 suppliers
search(2) → by meaning: "late delivery penalty liquidated damages"
grade     → 2 of 3 contracts answered; the third returned a
            general terms page
retry (2) → by exact words: "late delivery" in supplier 3's contract
          ← clause 9.2 found
answer    → 3 suppliers, each penalty clause cited to its page

What classic RAG would have done. Embedded the whole question once, fetched passages about deliveries and penalties from the contracts index, and never consulted the deliveries table — the only place “late last quarter” is recorded.

Research lineage

Agentic RAG vs RAG over time: the papers that moved retrieval.

Agentic RAG is a convergence of ideas rather than one invention: retrieval as memory, reasoning interleaved with action, models that judge their own evidence, and graph-shaped knowledge.

  1. May 2020RAGLewis et al. combine a seq2seq model with a dense vector index of Wikipedia (NeurIPS 2020).
  2. Oct 2022ReActYao et al. interleave reasoning traces with actions such as search calls — the agent loop that agentic RAG builds on.
  3. Oct 2023Self-RAGAsai et al. train a model to retrieve on demand and to critique the passages and its own output.
  4. Dec 2023Naive, advanced, modularGao et al.’s survey names the three RAG paradigms.
  5. Jan 2024Corrective RAGYan et al. add a retrieval evaluator whose confidence decides the next action, with web search to extend weak results.
  6. Apr 2024GraphRAGEdge et al. build an entity graph and community summaries for questions about a whole corpus.
  7. Jan 2025Agentic RAG surveySingh et al. define agentic RAG as embedding autonomous agents in the RAG pipeline.
Building blocks

The agentic RAG patterns: routing, rewriting, grading and teams.

The 2025 survey groups agentic RAG designs by four agentic patterns — reflection, planning, tool use and multi-agent collaboration. In a working system they show up as the parts below.

Agentic RAG agenta model with search tools and the judgement to use them
Routerpicks the source and method for each step (tool use)
Query plannerrewrites and splits the question (planning)
Retrieval toolsmeaning, exact words, filters, SQL, web
Graderjudges whether passages answer the question (reflection)
Corrective stepretries another way when the grade is low
Specialist agentsone per source, coordinated by a lead (multi-agent)
Citationsthe passages the answer actually relied on
Honest refusalsays the sources do not cover it

Each pattern, and the research that showed it working

  • Tool useReAct (2022) lets the model issue search actions between reasoning steps and read the results before the next step.
  • ReflectionSelf-RAG (2023) trains the model to emit reflection tokens that judge whether a passage is relevant and whether its own answer is supported.
  • CorrectionCorrective RAG (2024) puts a small evaluator after retrieval; its confidence decides whether to refine the passages, search the web, or both.
  • Planning and teamsThe 2025 survey catalogues designs where a planner splits the question and specialist agents each own one source.
Decision guide

Agentic RAG vs RAG: when the extra loop pays for itself.

Every grading step and retry is another model call, so agentic RAG costs more and answers more slowly. It earns that back on questions that span sources or need more than one hop.

Multi-step question
Deep questions on one libraryComparisons, “which of these…”, questions that need two searches. Add query planning and grading.Agentic RAG
Research across systemsDocuments, records and the web together. Routing and grading are the whole value.Agentic RAG
Single lookup
FAQ over a policy libraryOne index, short factual questions. A fixed pipeline is faster and cheaper.Classic RAG
Lookups across a few sourcesA router in front of fixed pipelines often suffices; add grading when wrong-source answers appear.Routed RAG
One sourceMany sources

Costs agentic RAG adds

  • LatencyEach plan, grade and retry is a model round trip.
  • TokensPassages are read by the grader and again by the writer.
  • VarianceTwo runs can take different paths; evaluations must test paths too.

What it buys

  • Right sourceNumbers come from tables, wording from documents.
  • Fewer confident errorsWeak retrieval is caught before it reaches the answer.
  • Honest gapsThe agent can say the sources do not cover a question.
On Botminds

How Botminds agents run agentic RAG on your knowledge.

Botminds agents answer from your libraries, documents, records and the web. Each one decides per question whether to search, where and which way — and every search runs under the asking person’s permissions.

More than two dozen ways to search

By meaning, by exact words, by label, by document or page, page and document summaries, records counted in the database, chat files and the web — often several for one answer.

A coverage check

A search whose passages probably do not answer the question comes back with a note, so the agent searches again or says so.

Knowledge lock

Choose to have the agent answer from its sources alone, and say plainly when they do not cover a question.

Permission-aware retrieval

The workspace, the libraries the agent was given, its view and the asker’s read rules apply inside every search.

Citations it used

The agent marks the passages it relied on; those are the sources shown, each with its document and page.

Failure said as failure

A failed search says it failed, and a library still indexing says how much of it is ready.

Read the full design in agentic RAG, and build one in the Academy lesson the search pattern.

Questions people ask

Agentic RAG vs RAG: common questions.

The technology behind itAgentic RAGEvery way to search, at any scale
What is agentic RAG vs RAG in simple terms?

In agentic RAG vs RAG, classic RAG searches once and answers from whatever came back. Agentic RAG puts an agent in charge of the search: it can skip it, run several, pick different sources and methods, and check the results before answering.

Is agentic RAG better than traditional RAG?

For questions that span several sources or need more than one search, usually yes. For short factual lookups over one library, classic RAG is faster and cheaper and often just as accurate. Measure both on your own question set.

What is naive RAG?

Naive RAG is the basic pipeline: chunk, embed and index documents, then for each question embed it, fetch the top-k nearest passages and generate an answer. Advanced RAG adds query rewriting and re-ranking; modular RAG makes each stage a swappable module.

What is the difference between agentic RAG and graph RAG?

They answer different questions. Graph RAG changes how knowledge is stored — an entity graph and community summaries — so whole-corpus questions can be answered. Agentic RAG changes who controls retrieval. An agent can use a graph as one of its tools.

What are corrective RAG and Self-RAG?

Both add self-checking. Self-RAG (2023) trains a model to retrieve only when needed and to critique passages and its own output. Corrective RAG (2024) scores the retrieved documents with a small evaluator: documents it judges correct are refined to their key strips, incorrect ones are replaced by a web search, and when it is unsure it combines both.

Is agentic RAG slower and more expensive?

Yes, per question. Planning, grading and retries each add a model call. The cost is worth it where an answer from the wrong source, or with no evidence behind it, costs more than a few seconds and a few cents.

Is agentic RAG the same as agentic AI?

Agentic RAG is one use of agentic AI: an agent whose tools are searches. Agentic AI in general also covers agents that take actions — updating records, sending messages, running workflows.

Does Botminds use agentic RAG?

Yes. Botminds agents decide per question whether and where to search, pick from more than two dozen ways to search, check coverage, run every search under the asker’s permissions and cite the passages they used. See agentic RAG.

Sources
  1. Lewis et al. (2020): Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasksarxiv.org
  2. Gao et al. (2023): Retrieval-Augmented Generation for Large Language Models — A Surveyarxiv.org
  3. Singh et al. (2025): Agentic Retrieval-Augmented Generation — A Survey on Agentic RAGarxiv.org
  4. Yao et al. (2022): ReAct, interleaving reasoning and acting in language modelsarxiv.org
  5. Asai et al. (2023): Self-RAG — Learning to Retrieve, Generate, and Critique through Self-Reflectionarxiv.org
  6. Yan et al. (2024): Corrective Retrieval Augmented Generationarxiv.org
  7. Edge et al. (2024): From Local to Global — A Graph RAG Approach to Query-Focused Summarizationarxiv.org
Read next
RAG vs fine-tuningWhich fixes missing knowledge, which fixes behaviour, and when to use bothKnowledge graph vs vector databaseRelationships against similarity: what each stores, answers and costsKeyword vs semantic searchBM25, embeddings, hybrid search and re-ranking, and where each one failsAll guidesEvery explainer, by topic

See it on
your own work.

Bring one process and its documents. We’ll show you the agents, the controls and the record, running on your data.

Book a demoAgentic RAG