pdf → document object → evidence

Convert PDFs into
evidence-backed
document objects.

Structured for LLM use: math, geometry, images and tables — not just plain text. Every result comes with a report showing exactly where to look or correct.

▤ open the evidence reports ⇄ translate a PDF ◳ browse a document model ⎇ source

01 — ingestion

Three routes in, one shape out.

Whatever the PDF, ingestion ends in lines.json — page geometry plus recognised text and math — so everything downstream works the same way.

Optimal OCR route

MathPix→lines.json

The best-quality route for math.

Alternative OCR

Gemma 4 26B→lines.json

Hosted model, MathPix-compatible output.

Binary PDFs

pdfminer.six→lines.json-like

Math-extended pdfminer.six reads the text layer directly.

02 — document object functions

Operate on the object, not the text.

Transclusion

Every math expression is held once and transcluded wherever it is used.

Translation

The transclusions map all non-textual elements to text that is not changed by an automatic translation via DeepL.com.

→ worked ZH → EN example

Bibliographic aggregation

References resolved and aggregated via Perplexity.ai search.

Conceptual summaries

A conceptual summary for each paragraph.

Spoken math

Formulas rendered as speakable text.

03 — output projection

One model, four projections.

Markdown

Readable text with math inline.

LaTeX

LaTeX source projected from the model.

LLMtext

Token-optimised math for LLMs.

TiddlyWiki JSON

Tiddlers as JSON, ready to import into TiddlyWiki.

→ example: kolbe2018hubbard.tiddlers.json

04 — evidence reports most important

See exactly where to look.

Each report compares what pdfdrill produced against the original page, so a reader — human or agent — checks evidence instead of trusting a result.

HTML inspect

The element tree, with each node's region highlighted on the page.

→ example: arXiv 2010.14265

HTML formula tables

LaTeX source · KaTeX render · MathPix crop, side by side for every formula.