Skip to content
Natural PDF

Solve: real PDFs, worked end to end

Every page in this section starts from a genuinely bad PDF — the kind that shows up after a records request or a scrape — states why it resists extraction, and works the solution in runnable code. Each page is also downloadable as a notebook. Same-repository examples use checked-out files under pdfs/ during the docs build, and notebook export rewrites those paths to raw GitHub URLs so you can still run them top to bottom in Colab. A few externally hosted records remain URL inputs because they are not committed to this repository.

If you’re new to the library, start with Learn; come here when you have a document that looks like one of these.

Tables without borders

Section titled “Tables without borders”
Oklahoma alcohol licensee list Zebra-stripe tables with no ruling lines An Oklahoma licensee list where alternating row colors stand in for borders and cells wrap across lines — the table detector merges everything. Vancouver use-of-force records Use-of-force records in six-point type Police use-of-force data in tiny type with acres of whitespace and not one ruling line — detectors see haze, not columns. Niagara Falls 911 call log A 43-page borderless call log A spreadsheet printed to PDF: erratic column widths, columns that go empty for pages, and a "2770 Records Found" line aimed at your data.

Tables that span pages

Section titled “Tables that span pages”
Serbian regulatory fee table A table that runs for eight pages of a Serbian law One fee table from page 90 to 97 with no boundaries between pages, section headers mid-stream, and a formula that's an image, not text.

Scans and OCR

Section titled “Scans and OCR”
Pixelated call center report A pixelated FOIA scan with an invisible table A call-center report scanned so badly the numbers are hard for humans — no text layer, half-dissolved ruling lines, OCR required.

Multi-column layouts

Section titled “Multi-column layouts”
Three-column page with two tables Multi-column pages that read in the wrong order Three columns with tables buried in them — naive extraction reads straight across all three at once, interleaving unrelated sentences.

Redactions and form printouts

Section titled “Redactions and form printouts”
Law enforcement complaint records A relational database printed to PDF, with redactions Complaint records as repeating form blocks with one-to-many tables inside — and redaction boxes that break column detection where it hurts.

Non-Latin and right-to-left scripts

Section titled “Non-Latin and right-to-left scripts”
Tunisian election results in Arabic Arabic election results across four pages A Tunisian results table in right-to-left Arabic script with spanning and rotated headers, split over four pages.

Where these come from

Section titled “Where these come from”

These documents are drawn from Bad PDFs, a collection of real problem PDFs submitted by journalists and researchers, each with a worked Natural PDF solution. If you have a PDF that deserves a page here, submit it there.