Arag

This entry covers an agentic retrieval augmented generation framework from researchers at the University of Science and Technology of China and Metastone Technology in Beijing. Instead of retrieving passages once or forcing a model through a fixed workflow, the framework exposes hierarchical retrieval interfaces directly to the model and lets the agent decide when to retrieve, what to search for and which tool to use. It is aimed at researchers and engineers building question answering and knowledge systems on top of frontier models with strong tool use. The framework provides three core tools: keyword search for exact lexical matching, semantic search for dense retrieval, and a chunk read tool for opening full document chunks. The agent combines them in any order and adapts its strategy to intermediate results. The accompanying preprint evaluates it against GraphRAG, HippoRAG2, LinearRAG and other baselines on multi hop question answering sets such as MuSiQue, HotpotQA and 2WikiMultiHop, a medical question set and a long context novel benchmark, and reports the strongest results among the methods compared. The code is published free on GitHub alongside the paper on arXiv. There is no commercial product, hosted service or pricing; it is a research release that users run and adapt themselves. The project is presented as a 2026 preprint on the lab's agent research site, so it should be treated as research code rather than a supported product, with updates tied to the authors' ongoing work.