AI/ rag · llm · ai-research · benchmarking

New RAG System Trims Graph Costs by More Than Half

A new preprint argues that graph-based RAG can be both accurate and cheap, undercutting a rival system's costs by more than half in early tests.

A new RAG system says the graph-building step that makes retrieval-augmented generation smarter doesn't have to blow up the bill.

The system, called EffiRAG, appears in an unreviewed arXiv preprint (arXiv:2609.18099); it has not yet been through peer review. EffiRAG uses a knowledge graph only to locate relevant passages, then pulls answers from the original source text instead of generating from the graph itself. Tested on UltraDomain, a 120-question benchmark spanning four domains, EffiRAG's answers were preferred over rival system LightRAG-hybrid on 93 questions, with LightRAG winning 7 and 20 rated a split. The preprint reports EffiRAG cut total system cost, covering every language-model call during ingestion and querying, by 57 percent, from $0.952 to $0.408 per run.

Graph-based RAG has a known problem: building the graph means burning language-model calls on every document before a user asks a single question, which makes it expensive to scale. The preprint's bigger claim is that cost should be reported alongside quality as a standard metric, since EffiRAG's edge held, and grew, as the corpus expanded. At 10 and 20 documents per domain it stayed preferred over LightRAG-hybrid while costing 4.2 times and 4.5 times less, helped by a non-LLM filter that skips low-value chunks before the graph gets involved.

That's one paper's benchmark against one rival, not an industry standard, and arXiv:2609.18099 hasn't been through peer review yet. Worth remembering before anyone rewrites a retrieval pipeline based on a preprint's own scorecard.

TR

The Revision

Written by an AI system from the public sources credited above. How we write →