Jev in production › Document and record classification

PDF Race

Benchmark app in which Jev labels Docling-parsed arXiv papers by category and picks each title from candidate lines, raced against two Gemini 3.8 Flash lanes on the same papers. The Jev lane cost $0.0022 in total (author).

Open on GitHub ↗

$0.0022measured against a baseline, as published by the source
Use
Document and record classification
Industry
Data and analytics
Form
Open-source tool
Stage
Beta
Listed
2026-09-22
Found via
github
Repository
goodrahstar/pdf-race
Stars
13
Forks
5
Last push
2026-09-22
Language
JavaScript
License
MIT

The README opens with

A parser and two models race to read documents: Docling → Jev (TypeSafe's decision model), Docling → Gemini 3.8 Flash, and Gemini 3.8 Flash reading the PDF itself. Every lane answers the same questions, and the answers are scored against arXiv's own metadata.

12 arXiv papers, 189 pages, 4 questions each. Every lane got 12/12 categories and 12/12 titles. The difference was speed and cost.

Lane Wall clock Decide only Cost Category Title --- ---: ---: ---: :---: :---: Docling → Jev 24.5 s 6.5 s $0.0022 12/12 12/12 Docling → Gemini 3.8 Flash 26.7 s 48.5 s $0.0364 12/12 12/12 Gemini 3.8 Flash reads the PDF 32.1 s 84.1 s $0.0882 12/12 12/12

Badge

For the project's own README, linking back here:

Listed in Jev in production

Also used for document and record classification