agentikus RAG

in English: agentic RAG (agentic retrieval-augmented generation)

advancedspreading

An AI solution where an agent itself decides what, where, and how many times to search sources before formulating an answer.

In plain terms

Traditional RAG works like a hurried student: it glances at the notes once for your question, highlights the first few results, and answers based on those. If it opened the notebook in the wrong place, the answer will be wrong too.

Agentic RAG is more like a good librarian. First, it thinks through what you are curious about, then looks around on several shelves. If the first result is insufficient or contradictory, it rephrases the search and tries again. It only answers when it has gathered enough material.

This is especially useful for complex questions that require comparing multiple documents or reaching the answer in multiple steps. The cost is that it is slower and more expensive than a single search.

How it works

The agent receives your question and breaks it into subtasks: what it needs to learn in order to answer.
It selects where it is worth searching: from an internal document repository, a database, the web, or another tool.
It runs the search, then evaluates the results: whether they are relevant to the question, fresh enough, and if anything is missing.
If there is insufficient material, it reformulates the search or turns to another source, repeating this until it has enough basis for the answer or reaches the set step limit.
Finally, the language model writes the answer from the collected text segments, ideally citing the sources.

An everyday example

You ask the corporate AI assistant whether the home office policy has changed since last year and how this affects your contract. The assistant retrieves the old and new policies, compares them, checks the section relevant to your contract type, and only then responds. You see the same in the "deep research" functions of chatbots, where the AI searches and reads for minutes before writing a summary.

Who is affected

employeeCorporate knowledge base search assistants increasingly work this way, so you get more useful answers to complex questions. → Work & HR

entrepreneur, executiveIt usually provides more accurate answers to complex questions based on your own documents, but this comes with higher computational costs and longer response times, which should be weighed in advance.

developerInstead of a simple search chain, you must design a decision cycle, tool usage, stop conditions, and thorough evaluation. → Code

lawyer, citizenIt is a useful aid for questions comparing many documents, but you must always verify the provided references yourself. → Law

A common misconception

MythIf the AI checks the sources multiple times, its answer is definitely correct.

In realityMulti-round search reduces the chance of error but does not eliminate it: the agent may misunderstand the question, select a poor source, or draw an incorrect conclusion from a good source.

Limits and risks

It is slower and more expensive than a single search because each additional cycle requires another model call.
It is only as good as its sources: it compiles convincing but incorrect answers from outdated or faulty documents.
An agent can get lost: it may run unnecessary loops or stop too early, requiring step limits and human oversight.
Instructions hidden in texts read during a search can deceive the agent (prompt injection), especially if it also searches the open web.

Similar, but not the same

RAG (retrieval-augmented generation) – Traditional RAG searches once in a fixed sequence and then answers, whereas in the agentic version the model itself decides how the search proceeds and whether to repeat it.
AI-agent – An AI agent can perform any task with tools, while agentic RAG specifically uses this capability for gathering information from sources and answering.
semantic search – Semantic search only returns results matching the meaning of the query; it does not formulate an answer. Agentic RAG uses this as one tool within a multi-step process.

Related concepts

Broader

RAG (retrieval-augmented generation), AI-agent, generative AI

Narrower

Multi-agent RAG, self-correcting RAG, deep research

On the edge

tool use, vector database, semantic search, hallucination, context window

A brief history

The concept of RAG was introduced in a 2020 paper by Patrick Lewis and co-authors. The agentic variant spread around 2023–2024, when language models learned to call tools reliably and to work in multiple steps.

Where things stand

As of late 2026, the research functions of major chatbots and many corporate knowledge assistants already work this way. However, measuring reliability and managing costs and security remain immature.

Status: 2026 November

Also known as

agentic RAG · agentikus RAG · agentic retrieval-augmented generation · agentic retrieval · agentic retrieval · agent-based RAG · agent-based RAG · agentic search

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.