AI glossary · Advanced · Agents · Language models
Agentic RAG
Agentic RAG is a form of RAG in which an AI agent plans and adjusts the search process itself. The AI decides what to look up again and when, so instead of searching once and stopping, it searches several times and refines the results before answering.
About this term
- Level
- Advanced · Specialist terms that need the inner workings
- Category
- Perception & action
- Fields
- Agents, Language models
- Added
- 2026-10-07
What is different
RAG (retrieval-augmented generation) is a technique that has a large language model (LLM) find relevant content in external documents or databases and refer to it before answering. This lets it use up-to-date information or internal company material that is not in its training data. However, traditional RAG follows a fixed sequence: 'when a question comes in, search once, attach a set number of documents and answer.' This makes it weak on complex questions that require reasoning through several steps. Agentic RAG puts an AI agent that makes its own decisions into this flow, so the search method and order can change to fit the situation.
How it works
The agent first looks at the question and plans whether a search is needed and what to look for and where. After reading the search results, it checks on its own whether the content is sufficient and relevant to the question; this is called 'reflection'. If something is missing, it changes the search terms or uses other tools (web search, database queries and so on) to search again. For complex work, multiple agents may split up the roles. It repeats 'plan → search → check → search again' in this way, and when it judges that the evidence is sufficient, it produces the final answer.
How it has developed
According to Wikipedia, the term 'RAG' was first used in a 2020 paper. The paper proposed having a language model search an external memory store and use it when answering. 'Self-RAG', published in 2023 by Akari Asai et al., trained a model to search only when retrieval is actually needed and to evaluate the retrieved documents and its own answers. It was an important step from always searching to 'the AI deciding for itself when to search'. The 2025 survey paper 'Agentic Retrieval-Augmented Generation: A Survey on Agentic RAG' by Aditi Singh et al. collected various approaches that combine agent techniques such as planning, tool use and multi-agent collaboration with RAG, and classified them by factors such as the number of agents, control structure and degree of autonomy.
Why it is getting attention
Questions that come up in real work are often hard to answer by finding just one document. Because agentic RAG can move across multiple sources and reason step by step, it can handle such questions better. The survey paper mentioned above compiled use cases in fields such as healthcare, finance, education and enterprise document processing. RAG can also show the sources of an answer, making it easier for users to check the content themselves, and the agentic approach is developing in a direction that keeps this advantage while improving search quality.
Things to know
Because it repeats search and checking several times, it can be slower and cost more than traditional RAG. If the agent makes poor decisions, it may keep finding irrelevant material or take unnecessary steps. As Wikipedia points out, the RAG limitation that the model can still produce wrong answers (hallucinations) by misreading the context, even when it retrieves the right material, also remains. The survey paper named evaluation methods, coordination among multiple agents, memory management, efficiency, and management and control (governance) as challenges still to be solved.
Example
Imagine asking an internal company AI assistant, "Did any of our team's projects go over budget last quarter, and why?" Traditional RAG might find a few similar documents and try to answer in one go, missing some information. An assistant using agentic RAG breaks the work into steps: it first finds the project list, checks each project's budget documents, and then searches meeting notes again for the projects that went over budget to find the reasons. Finally, it checks whether the evidence it found sufficiently answers the question and shows the documents it consulted as sources.