Reproduce “FuseRAG: Fusing Structure and Evidence into Inference for Cost-Efficient Graph-based RAG” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Tingqi Wang, Yingshu Li, Zhipeng Cai
FuseRAG uses graph retrieval-augmented generation by keeping evidence in reusable token-level latent states, adding query-conditioned graph interactions, and retaining selected facts as text for grounding. On 2WikiMultiHopQA with Llama-3-8B, its 128-token explicit-evidence budget achieves 55.35% F1, versus 36.53% for Clue-RAG-1.0, while TTFT is 62.25 milliseconds versus 132.55 milliseconds for RAPTOR. These results show a favorable answer-quality/first-token-delay trade-off on evaluated benchmarks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.