Building Document Structure with Loop Engineering: Recovering a PDF’s Outline from Body Typography for RAG

A detailed walkthrough of recovering a PDF's table of contents from body typography when no native outline or printed contents page exists, as part of an enterprise RAG pipeline. Six deterministic per-line signals (font size ratio, boldness, numeric prefix, short length, left alignment, blank line above) score heading candidates from an enriched line_df/span_df. A bounded LLM validation loop then filters false positives. The approach handles three scenarios: no TOC at all, partial native outlines needing deeper levels, and composite concatenated documents. Evaluated against six real PDFs, the deterministic pass achieves 72–100% recall on decimal-numbered documents but poor precision (7–38%), which the LLM validation lifts dramatically (to 72–100%) while preserving recall on most fixtures. The article also introduces a two-layer enrichment concept where typography-derived hierarchy (level) is complemented by business-domain tags per paragraph for more precise RAG retrieval.

Nguồn: https://towardsdatascience.com/building-document-structure-with-loop-engineering-recovering-a-pdfs-outline-from-body-typography-for-rag. 8sync News chỉ tóm tắt và dẫn link; bản quyền nội dung thuộc tác giả và nguồn gốc.