Agentic RAG

An evolution of traditional RAG (a linear pipeline: search → inject → answer). In agentic RAG, retrieval becomes one tool among many, and the model has the autonomy to plan, refine its search, evaluate what it retrieved, and decide whether to search again or whether it already has enough context.

The line between them: if there's a loop where the model decides the next step, it's agentic; if it's a fixed sequence, it's a pipeline. See Everyone's an agent now. Or are they?.