Skip to content
BENIBIX Learn

Agentic RAG

Also known as Agentic Retrieval, Multi-Step RAG

RAG where the model does its own search, a few times, then answers.

Think of it like

Think of it like a detective. They look for clues, read them, then pick where to look next.

Example

Emma asks which of three suppliers raised prices most since 2024. The model searches each supplier's price list in turn, then compares them and answers.

Why it matters

Many real questions need more than one search. Agentic RAG can follow the trail.

Definition

Agentic RAG lets the model run the search itself, in as many rounds as it needs. Instead of one fixed search before the answer, the model writes its own queries, reads the results and decides what to look up next. It can search several sources and split a hard question into parts. It stops when it judges it has enough to answer. That handles harder questions, but it costs more time and tokens per answer.
The model searches, reads and checks if it has enough. If not, it goes round again.

Common mistake

Using it for simple lookups. One search is faster and cheaper when one search is enough.

Key takeaways

  1. 1
    The model searches. It writes the queries.
  2. 2
    Several rounds. Read, then search again.
  3. 3
    Costs more. Use it for hard questions.