01
A PDF is not a document
What the file actually contains, the three things it does not, and why every word and table has to be inferred from geometry.
3 Aug 20263 min
Engineering notes
A PDF is not a document.
It is a set of printer instructions.
So to read one, you have to think like a printer.
Notes from building the retrieval engine — what we measured, what broke quietly, and what the numbers changed our minds about.
- Word drift
- 11.99pt→0.0000pt
- Signs kept
- 81%→100%
- Fonts covered
- 1→12
- Table F1 ceiling
- 0.362→0.935
Reading a PDF like a printer · 7 parts
- 025 min
Thinking like a printer
Advance widths, the fourteen fonts every reader is assumed to know, descenders, and the three namespaces between a byte and a character.
- 034 min
Where tables actually live
A table is never stored — it is reconstructed. Two strategies, two failure modes, and the geometry that makes citations clickable.
- 044 min
Building pdftable
The architecture, the API, and why matching pdfplumber exactly was a better decision than improving on it.
- 055 min
The bugs that never crashed
Five silent corruptions, including eight thousandths of a point that flipped the sign of a number in an annual report.
- 066 min
Measuring instead of believing
Benchmarks, a harness that measured itself twice, and the number that redirected the whole project.
- 075 min
What is still missing
The gaps, the models worth deploying, the ones that are not, and three lessons that are not PDF-specific.