A toolkit for scientists who refuse to flatten their PDFs

What it does

PaperMage ingests academic PDFs and exposes their contents through a Document class that carries multiple segmentation layers—tokens, sentences, rows, pages, figures, tables, and more. Instead of forcing these layers into a rigid hierarchy, it treats them as intersecting sets of Entity objects backed by spans and bounding boxes. The library wraps existing utilities like PDFPlumber and PDF2Image, then stitches their outputs together with in-house predictors that can be heuristics or full machine-learning models.

The interesting bit

The real trick is the dynamic indexing: every time you add a new layer, the document rebuilds intersection maps so any Entity can instantly retrieve whatever else overlaps with it. This means a sentence can ask for its rows even when those rows bleed past the sentence boundary, which is exactly the kind of mess real scientific PDFs serve up.

Key highlights

• Document.symbols holds the raw text, while named layers (pages, tokens, sections, etc.) provide alternate slices into the same string.
• Entities expose both .spans (character ranges) and .boxes (page coordinates), linking textual and visual representations.
• Predictors can be stacked incrementally: load a saved document, run a new model, save the enriched JSON, and repeat.
• Ships with a CoreRecipe that bundles parser, rasterizer, and default predictors into a single pipeline.

Caveats

• The authors explicitly warn this is a research prototype for EMNLP 2023 and regular maintenance is unlikely; active development has moved to the Dolma project.
• Intersection logic is non-obvious: because layers overlap rather than nest, iterating sent.rows can yield tokens that sit outside the sentence entirely.

Verdict

Worth a look if you are building multimodal scientific-NLP prototypes and need a sane abstraction over PDF chaos. Skip it if you want a production-grade, maintained parser—head to Dolma instead.