Chapter 5 of 7
Abracadabra in 5 minutes.
Drop a file and get on with your day. The bucket does its tricks in the background and tells you when your set is ready.
About those 5 minutes Target
Five minutes for a typical document is our design target, not yet a measured promise. We'll publish measured times after the first pilots. A long, multi-page bundle takes longer than a one-page invoice.
What happens in those minutes
- Prepare. The file type is recognized by its content. A PDF becomes page images plus its text layer, and every page gets a readability test: blur, resolution, blank page.
- Recognize. Two models look at the first pages and name the document type; if they disagree, a stronger model decides. The type must have an approved template.
- Read. Three models from three vendors, two passes each, all in parallel. For every field: the value, the page and the quote, or an honest "unreadable" or "not in document".
- Vote. Normalize the readings, apply the agreement rule, run check digits and cross-checks.
- Arbiter. A stronger model takes a second look at disputed fields only.
- Summary. One model writes the short summary using nothing but the document; a model from another vendor checks every sentence against it. Unconfirmed sentences are removed and noted in the receipt.
- Deliver. The full set lands in storage with its status, and you get notified. Every step has a fallback, so a set with a receipt is always produced.
Fish it out your way
- download it in the browser
- poll the API
- subscribe to an event notification (Amazon SNS)
- or let your IT team pick up sets straight from storage (Amazon S3)
Blurry scan? We don't waste spells
Pages that can't be read are caught before any model runs. A document with no readable page ends as unreadable right away, at no reading cost. A monthly page limit for each organization keeps the bill predictable.