Agentic RAG

Agentic RAG hands control of the retrieval process to the model itself.

Plain RAG always retrieves once and always with the user's question.

An agent decides whether to search at all, since some questions need no documents.

It chooses what to search for, which may be nothing like the original phrasing.

It can read the results and decide they are insufficient.

It can then search again with a better query, or search a different source.

It can break a multi-part question into parts and handle each separately.

It can compare what it found against the question before committing to an answer.

That self-correction is the real gain, since plain RAG cannot recover from a bad retrieval.

Corrective RAG formalises it by grading retrieved documents and re-searching when they score badly.

Self-RAG adds tokens that let the model decide when retrieval is needed.

The cost is several model calls per question instead of one.

Latency goes from a second or two to five or ten.

It needs a hard cap on iterations, or a confused agent will search indefinitely.

Use it for genuinely complex questions and keep plain RAG for the simple majority.