A new retrieval-augmented generation system called Proxy-Pointer RAG claims to achieve 100% accuracy on complex financial documents by preserving structural information that standard vector systems discard. The approach, detailed in a recent article, processes structured documents like 10-K filings by maintaining their hierarchical organization rather than shredding them into flat chunks.

The system was stress-tested on four major companies' annual reports - AMD (121 pages), American Express (260 pages), Boeing (190 pages), and PepsiCo (500 pages) - using 66 questions across two benchmarks. These included adversarial queries designed to break naive retrieval systems, such as calculating reinvestment rates from cash flow statements or analyzing multi-hop numerical relationships across financial statements.

Unlike conventional vector RAG that loses document structure, Proxy-Pointer injects breadcrumb paths into chunks and uses a two-stage retrieval process combining semantic search with LLM re-ranking. The complete pipeline has been open-sourced, allowing developers to test it on their own documents. The approach could fundamentally change how enterprises handle structured documents for AI applications.