Reproduce “TERAG: Token-Efficient Graph-Based Retrieval-Augmented Generation” on OpenResearch

About this paper

Qiao Xiao, Hong Ting Tsang, Jiaxin Bai

TERAG introduces a token-efficient framework for graph-based Retrieval-Augmented Generation (RAG) that significantly reduces knowledge graph construction costs by minimizing LLM output token consumption by 89-97%. This approach maintains competitive retrieval accuracy on multi-hop QA datasets, achieving or exceeding 80% of strong baselines while being substantially more economical.

Key insights

• Minimizing LLM involvement to direct, single-pass extraction of core semantic units is highly effective for drastically reducing output token consumption during knowledge graph indexing.
• A simplified graph structure, combined with efficient non-LLM methods for defining edges and employing a Personalized PageRank for traversal, can effectively support multi-hop reasoning.
• The design of the PPR restart vector, which integrates both semantic relevance and concept frequency, consistently improves retrieval accuracy by providing a more discriminative initial distribution.

Summary by alphaXiv.