DocuBench pairs 72 real-world documents with a JSON Schema and a hand-verified label, then scores each system on macro-average field accuracy. The set is deliberately hard: arrays and nested tables, right-to-left, CJK, and Devanagari scripts, rotated scans, handwriting, and non-PDF formats. Results as of July 31, 2026.
DocuPipe built DocuBench and holds its own results to the same bar as everyone else: every system gets the same schemas, the same hand-verified labels and the same open scorer, and every raw output is published, so anyone can reproduce these scores. It measures field extraction against a schema, not OCR text accuracy or chunking for search. Code is MIT licensed; each document keeps its original license, listed in SOURCES.md.
Methodology, the scoring contract, worked examples, and a per-document analysis of where each system wins and loses.
The DocuPipe rows above are the same extraction product you can use today: define a schema, upload your documents, and get structured JSON in the dashboard or through the API.
Practical guides to the two most common jobs behind these numbers: extracting invoice data for accounts payable, and extracting structured data from PDFs in general.