allenai/papermage

Provides multi-layered document parsing that segments PDFs into dynamically-indexed entity types (pages, tokens, sentences, sections, figures, tables, etc.), enabling cross-layer queries where entities can be traversed regardless of nesting boundaries. Built on modular `Parser`/`Rasterizer`/`Predictor` components—wrapping PDFPlumber for text extraction, PDF2Image for rendering, and custom ML models for semantic segmentation. Documents serialize to JSON for reproducible pipelines combining heuristic and learned extraction.

791 stars and 313 monthly downloads. No commits in the last 6 months. Available on PyPI.