Agentic RAG

Agentic RAG upgrades ordinary retrieval-augmented generation from a fixed 'retrieve once, then answer' pipeline into an agent that decides when, what, and how many times to retrieve. Instead of always pulling the top chunks for the raw question, the agent can reformulate the query, search multiple sources, judge whether the results are sufficient, retrieve again with a better query, and only then answer — treating the knowledge base as a tool it calls in a loop. This handles multi-hop questions ('which of our enterprise customers churned after a price change?') and vague queries that plain RAG answers poorly. For builders, agentic RAG noticeably improves answer quality and reduces 'the docs didn't have it' misses, at the cost of more model calls and latency per question. Start with simple RAG, measure where it fails, and add agentic retrieval only for the query types that actually need multi-step lookup rather than everywhere.