Reproduce “Evidence Units: Ontology-Grounded Document Organization for Parser-Independent Retrieval” 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

Yeonjee Han

The paper introduces "Evidence Units" (EUs) for document chunking in Retrieval-Augmented Generation (RAG) systems, designed to group semantically complete information and operate independently of specific document parsing outputs. This approach enhances retrieval performance, achieving a 61% increase in Avg LCS and a 3.4-fold improvement in Recall@1 on the ground truth evaluation track.

Key insights

  • By proactively grouping semantically cohesive elements into "Evidence Units," the coherence of retrieved information for RAG systems significantly improves, enabling more complete answers from single retrievals.
  • The aggregated spatial footprint of Evidence Units converges across different parsers, effectively abstracting away the variability of individual element bounding boxes and labels to achieve parser-independent chunking.
  • Formalizing chunking rules and quality assurance invariants within a knowledge graph provides a flexible, maintainable, and verifiable framework for document organization that adapts dynamically.

Summary by alphaXiv.

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