Agentic RAG

Agentic RAG studies systems where a language model actively searches, inspects evidence, and decides whether to retrieve again before answering, rather than relying on a single retrieval pass. Common approaches include query decomposition, tool-using or multi-agent workflows, graph and hierarchical retrieval, and reinforcement learning that rewards useful evidence gathering and efficient stopping. These methods are tested on multi-hop question answering, long and multimodal documents, and domain-specific tasks such as clinical, legal, and financial reasoning. A central challenge is balancing answer quality with retrieval cost and latency while ensuring that evidence is sufficient, faithfully used, and not undermined by errors that accumulate across steps.