Reproduce “LivingRAG: Augmenting Graph RAG with Experience” 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

Yuzhuo Cui, Zongye Zhang, Qingjie Liu

LivingRAG enhances graph-based Retrieval-Augmented Generation by creating a writable experience store that accumulates and reuses verified reasoning from past queries. The system improves QA accuracy on multi-hop benchmarks while reducing LLM completion tokens by 22.7% on average, leading to lower estimated API costs.

Key insights

  • Online Question Answering streams inherently contain significant reusable signals, such as shared entities, graph neighborhoods, and reasoning templates, which can be leveraged for efficiency.
  • Storing and reusing verified reasoning experiences, composed of sparse graph activation maps and compact reasoning summaries, effectively guides both information retrieval and LLM generation.
  • Rigorous quality gates for novelty and factual grounding are crucial to ensure the accumulated experience is reliable and prevents degradation of system performance.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.