Agentic retrieval is the broad pattern of embedding vector search inside an autonomous agent’s reasoning loop, rather than treating it as a one-shot lookup. The agent decides when retrieval is useful, formulates the query, inspects the results, and may act on them by retrieving again, calling another tool, or answering directly.
This reframes the vector database as an active participant in reasoning — a queryable memory the agent consults on demand — rather than a passive store hit once per request. An agent answering a support question, for instance, might first retrieve the user’s account history, then retrieve relevant documentation based on what it found, chaining searches together.
Agentic retrieval is closely related to agentic RAG; the distinction is mostly emphasis. Agentic RAG focuses on grounding generated text in retrieved evidence, while agentic retrieval describes the more general capability of an agent to use search as one of several reasoning tools.