A new paper claims a leaner way to do graph-based retrieval-augmented generation can cut costs by over 99% without sacrificing answer quality.
Researchers describe LiteRAG, a graph-based retrieval method that swaps out expensive retrieval-time LLM control for algorithmic exploration guided by the query, plus a reasoning-chain approach to building context. Tested on DistComp, a benchmark for multi-hop question answering over distributed-systems papers, LiteRAG reportedly posts the highest overall quality score (0.798) among the methods compared, while cutting per-query latency by more than 100x and cost by more than 99% versus GraphRAG Global and DRIFT. On a second benchmark, UltraDomain, the paper says LiteRAG matches LinearRAG's quality using about 14x fewer tokens. The authors credit two techniques, query-adaptive thresholding and community-aware hub penalization, for most of the token savings, according to an ablation study in the paper.
Multi-hop question answering, where a model stitches together facts from several documents, is one of the pricier things to do with retrieval-augmented generation because graph traversal and LLM-based ranking pile up costs fast. If LiteRAG's numbers hold up outside the paper's own benchmarks, that is the difference between a RAG system that is a research toy and one a company can run on real support tickets or code search without a server bill that scales with usage.
Any paper that introduces its own benchmark and grades its own homework deserves a raised eyebrow, and DistComp and UltraDomain look built around the shape of LiteRAG's improvements, so the real test is whether outside teams see the same gains on their own data.