How it works
Below is everything that happens between your POST and your JSON, in the order it happens, including the parts that only matter when something goes wrong.
In plain English, the way you would brief a new clerk. "The name, NIC number and date of birth from these cards." No coordinate boxes, no sample set, no layout to register. That sentence becomes the extractor, and the extractor gets its own endpoint and key.
One multipart request, or a JSON body with an image URL. The key sits in the path, so there is no header to sign and no SDK to install. A phone photo held at an angle is fine.
curl -X POST "https://vision.olee.cc/api/v1/extract/YOUR_KEY" \ -F "image=@nic.jpg"
Hashed with SHA-256 and matched against the stored hash. We never hold the key itself, so a database dump cannot call your endpoint. A deleted extractor is dead immediately, whatever its status says.
ext_9f3a··············8c1d5a09f4b2···Taken inside a row lock, so two calls arriving at the same instant can never both claim the last read. Your allowance is spent first, then any extra reads you bought. If anything after this fails on our side, the read is handed straight back.
Rotated upright from its EXIF orientation, longest side brought down to 2000px, re-encoded as JPEG. Big enough to read a worn NIC, small enough not to spend your money on pixels nobody needs.
The image goes to the model with your extractor prompt attached. It sees the document, not a transcript of it, so position still means something: on a trilingual card it can tell the Sinhala name from the Tamil one because it can see where each sits.
Named fields, each with a confidence number. A value read cleanly scores high. A value derived rather than read, like a name transliterated into a script the card does not print, is capped lower on purpose. A character nobody could read comes back null, never as a guess.
"name_en": { "value": "DILSHAN ABEYGUNAWARDANA", "confidence": 0.94 }
"name_sn": { "value": "දිල්ෂාන් අබේගුණවර්ධන", "confidence": 0.91 }
"nic_default": { "value": "200000112345", "confidence": 0.97 }
"date_of_birth": { "value": "2000-01-01", "confidence": 0.95 }A signed webhook to your own endpoint if you turned one on. A row in your Reads inbox with the image and the extracted data if you turned storage on, and nothing but status, timing and cost if you did not.
If we cannot complete a read, the read is returned to your allowance. You are charged for answers, not attempts.
If the document genuinely has nothing on it, that is an answer. We processed it and paid for it.
An unreadable digit comes back null with confidence 0. On an identity document a plausible wrong number is far worse than a gap you can see.
Your documents are used to answer your call. That is all.
The playground runs the real pipeline, so what you see there is what your code will get.