Reproduce “GraFine: Retrieval-Time Refinement for Efficient Graph RAG over Corpus Graphs” 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

Seonho An, Chaejeong Hyun, Min-Soo Kim

GraFine, developed at KAIST, introduces an efficient framework for retrieval-time refinement in Graph RAG over corpus graphs. The method achieves superior retrieval accuracy and RAG generation quality by integrating semantic and topological information, leading to average gains of +9.9% in Recall@10 and +9.1% in nDCG@10 while being significantly faster than LLM-interleaving methods.

Key insights

  • Existing graph retrieval operators (O_g and O_m) have fundamental 'blind spots' concerning semantic relevance during expansion and structural context during pruning.
  • Combining semantic signals with graph topology in expansion (O_vg) and incorporating graph context into model-based pruning (O_gm) leads to more effective and efficient retrieval.
  • The newly introduced Topological Recall (TR) metric, especially its MissTR component, effectively quantifies the refinement potential and predicts the discovery of unretrieved oracle nodes.

Summary by alphaXiv.

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