agentic RAG

Plain RAG always retrieves once and then answers, on a fixed track. Agentic RAG hands that decision to the model itself: it decides whether to retrieve at all, what to search for, whether the results it got are good enough, and whether to search again — treating retrieval as a tool it can choose to call, like a researcher deciding when to go back to the library.

Retrieval becomes one tool among several that the agent can invoke inside a reasoning loop. The model can reformulate its own queries, hit different sources, critique what it pulled back, and stop only when it is satisfied. This flexibility handles open-ended, multi-step questions that a fixed pipeline cannot, but the price is more latency, higher cost, and less predictability — plus a new failure mode where the agent loops forever or stubbornly chases the wrong source.