In standard RAG, retrieval is a fixed step — the system always queries a knowledge base before generating a response. Agentic RAG makes retrieval a dynamic decision. The agent evaluates whether retrieval is needed, decides what to search for, refines the search based on initial results, and determines when it has gathered enough information to respond.
A finance agent asked to summarize credit exposure for a specific client might run one search for recent transactions, another for the client's contract terms, and a third for current credit policy — then synthesize all three into a coherent response. That is retrieval driven by reasoning, not a mechanical trigger.
