Retrieval‑Augmented Generation (RAG) reduces hallucinations of large language models (LLMs) by grounding them in external knowledge, yet conventional dense retrieval struggles with the multi‑step reasoning required by multi‑hop question answering. Graph‑based RAG can model such relationships but suffers from severe semantic drift and high online latency due to noisy global graph traversals. ISO‑RAG (Isoperimetric Retrieval‑Augmented Generation) addresses this by imposing a geometric constraint on the underlying knowledge graph. The graph is embedded into a hyperbolic Poincaré ball and node‑wise isoperimetric profiles are precomputed; during retrieval spurious edges are pruned, restricting the search to a strictly localized subgraph. This topological purification directly regulates Personalized PageRank (PPR) diffusion, ensuring exact and low‑latency convergence within the subgraph. Benchmarks on multi‑hop QA datasets show ISO‑RAG improves retrieval recall by ~10.0% and downstream exact match by 4.3%, achieving a superior accuracy‑efficiency trade‑off and eliminating the latency bottleneck of global traversals. Source code: https://github.com/ZaiizaiZHANG/ISO-RAG.
Review