papermage

Why we included this project

If you work with scientific PDFs and need to get from raw pages to something a model can actually use, papermage handles the messy middle. It parses a PDF into a single Document object with multiple segmentations, pages, rows, tokens, sentences, sections, figures, tables, and equations, all cross-referenced so you can move from a sentence to its rows or a page to its captions. That layered structure is what saves you from writing your own extraction and alignment code; parsers, rasterizers, and ML predictors all share the same data model. The recipe-based API makes it quick to pick up for prototyping attributed QA, information extraction, or layout-aware work on scholarly papers. One caveat worth knowing: this is a research prototype from EMNLP 2023 and is not actively maintained, so treat it as a reference implementation and data model rather than a production dependency.