Proxy-Pointer RAG: Structure Meets Scale at 100% Accuracy with Smarter Retrieval

Proxy-Pointer RAG is a structure-aware retrieval architecture that exploits document section headings to dramatically improve RAG accuracy. The approach uses five techniques: skeleton tree parsing, breadcrumb injection, structure-guided chunking, noise filtering, and pointer-based context loading. Benchmarked on 66 questions across four Fortune 500 FY2022 10-K filings (AMD, AMEX, Boeing, PepsiCo), it achieved 100% accuracy at k=5 and 93.9% at k=3. Key refinements include an LLM-powered noise filter replacing hardcoded rules, and a two-stage retrieval pipeline combining FAISS broad recall with an LLM structural re-ranker. The full pipeline is open-sourced under MIT license, runs on a single Gemini API key with no GPU required, and includes pre-extracted documents and benchmarking scripts.