A new open-source project applies Proxy-Pointer RAG to enterprise document comparison, tackling a genuinely hard problem: meaning in contracts and research papers rarely lives in isolated paragraphs. Instead it's spread across sections, hierarchies, and clauses spanning hundreds of pages. The architecture pairs hierarchical breadcrumb embeddings with a lightweight LLM re-ranker to retrieve and align scattered sections before any comparison happens.

The prototype processes credit agreements from Emerson (136 pages) and Texas Roadhouse (190 pages), plus academic papers on text-to-vector graphics. Comparison criteria like collateral structures or financial covenants trigger a two-stage retrieval pipeline that matches semantically aligned regions across documents. Reports include discrepancy ratings, risk directions, and persona-driven analysis tailored to legal or research contexts.

The full code sits in a GitHub repository with a five-minute quickstart. By decoupling the comparison engine from upstream extraction and downstream report formatting, the system adapts to new document domains without touching its core retrieval pipeline.