Plate I · Document extraction
HyperStract
A thousand-page book goes in as one file and comes back as text, because every page is read by a different machine at the same time.
In service https://hyperstract.com
A scanned book is not one job. It is nine hundred and sixty-six jobs that happen to be bound together. HyperStract cuts the file into pages, hands each page to a different card in the plant, and assembles the text as the answers come back.
The largest run to date is a nine hundred and sixty-six page ophthalmology textbook. It finished in seven minutes and thirty seconds, from upload to assembled text. On a single machine the same book takes the better part of a day.
What it reads
Documents go to a vision model that transcribes the page as it stands: columns, tables, footnotes, plate captions and the numbers inside them. Nothing is summarised and nothing is rewritten. When a page comes back thin, the page is sent again at a higher resolution, and the two readings are compared before either is kept.
Images sent through the API take a second model that accepts an instruction, so a caller can ask for a form to come back as fields rather than as prose.
The specification
- Input: PDF, single images, or a batch of images through the API.
- Output: plain text and per-page records, held in object storage and in a database, addressable per page.
- Scale: the fleet reads as many pages at once as there are free cards.
- Cost: the target for the nine hundred and sixty-six page book is under three dollars of electricity and machine time. The same book through a commercial interface runs about twelve.
- Interface: a browser for a person, a chat-completions endpoint for a program.
Why it is ours
No vendor sells a reader that runs on cards of this age. The inference server is built from source, the scheduler that deals pages to cards is ours, and the quality pass that catches a bad page is ours. That is the reason the price per page is what it is.