Agentic-RAG: Autonomous Information Foraging in High-Dimensional Vector Spaces

ABSTRACT

Standard Retrieval-Augmented Generation (RAG) relies on static, single-pass semantic searches that often fail on complex, multi-hop queries. We introduce Agentic-RAG, an autonomous framework where a specialized search agent actively 'forages' for information by dynamically rewriting queries, navigating knowledge graphs, and evaluating the relevance of retrieved chunks in real-time. By granting the retriever agent the autonomy to execute iterative search strategies and decide when sufficient context has been gathered, our model achieves a 47% accuracy improvement on the MuSiQue and HotpotQA benchmarks.