Glossary
Agentic RAG: search controlled by an AI agent
In short
Agentic RAG is a setup in which an AI agent decides whether to search, where to search and whether the found material is enough, and searches again when needed.
Search and RAG · agentic RAG
MMOOZZGG.com editorsChecked: 1 min read
For a child
You are making a school report about penguins. First you read an encyclopedia and see that it says too little about what they eat. So you get another book, then ask your teacher and check that everything fits. You decide where to go next. Agentic RAG is like such an independent student: it keeps searching until it has the whole answer.
In plain words
In a regular setup, the search through documents runs once and always the same way. In the agentic setup, an AI agent controls it. It reads the question and decides whether to search at all, in which database and with which query. After it looks at the result, it may rephrase the query, check another source, call a calculator or the CRM, and only then answer. For example, to the question “why did sales fall in Kazan”, the agent takes the figures from a report, then searches the correspondence for complaints from customers in that city and puts it all together.
For business
The agentic approach pays off on complex questions where the answer comes from several systems: sales analytics, complaint review, a briefing before negotiations. Each answer costs more and is harder to predict: the agent takes several steps, and each step is a model call. So teams set limits on the number of steps and a token budget per request. There is also a security risk: an agent with access to tools may read too much or take an action nobody expected. Give it minimal rights and log every action. The example of an AI department shows how to split such tasks among several agents.
For a specialist
Agentic RAG builds autonomous agents into the retrieval pipeline. According to the survey by Singh et al. (2025), agents rely on four patterns: reflection, planning, tool use and multi-agent collaboration. This removes the limit of a static pipeline, which lacks the flexibility for multi-step reasoning. The authors propose a taxonomy by the number of agents, control structure, degree of autonomy and the way knowledge is represented. A close early approach is Self-RAG (Asai et al., 2023), where the model decides when to retrieve fragments and rates what it found and its own output with reflection tokens. In practice, the loop is built on function calling: search is a tool, and the model chooses the arguments and the number of iterations. Besides answer quality, teams measure the number of steps, the cost and the share of loops that never end.
Example
A manager asks the assistant whether a customer can get deferred payment. The agent checks the debt in 1C, the accounting system, finds the payment terms in the contract, reads the internal rule on deferrals and only then answers.
Related terms
Questions and answers
Sources
Where companies use it
Not sure whether your company needs this?
AI solutions map