Read receipts, invoices and your own documents into data you can check
Describe the fields you need, send a photo or a PDF, and get JSON back. Values come with where they were printed, how sure ReceViz is, and which checks they passed, so your code knows what to trust and what to show a person.
How it works
Define the fields
Start from a receipt or invoice template, or write your own document type: each field's type, the labels it is printed under, and the checks it must pass. Publish it as a version.
{ "key": "amount", "type": "currency", "required": true, "semantic": "total_amount", "aliases": ["TOTAL", "AMOUNT DUE"], "confidence_threshold": 0.8 }Send a document
A photo, a PDF, or the text your app already read on the device, to one endpoint. Small documents come back in the same response; large ones by polling or a signed webhook.
curl "$RECEVIZ_API/extractions" \ -H "Authorization: Bearer $RECEVIZ_SECRET_KEY" \ -F "file=@receipt.jpg" \ -F "document_type=payment_receipt"Get checked data
The values, shaped like your document type and ready to store; the evidence for each one; and a clear answer to whether a person should look before you use them.
"review": { "required": false, "reasons": [] }
What makes it different
The evidence behind each value
Values come back with the page, the box and the printed text they were read from, a confidence from 0 to 1, and the checks they passed. Whoever reviews a result sees what the machine saw.
Checks first, a second reader when you ask
Patterns, ranges, dates and the document's own sums are checked on every standard request. VERIFIED mode adds a language model that re-reads only the uncertain fields; its answer must be printed on the lines it cites, and a disagreement is reported for review, never applied silently.
Long receipts keep their small print
Adaptive Document Tiling reads a long receipt or a large scan in overlapping regions at full resolution, then puts the lines back together in the page's coordinates, each line once.
Capture that keeps frames on the device
The Web SDK finds the document in the camera view, waits for a sharp, still, well-lit moment and takes one picture. Frames are analysed in the page and never uploaded; only the picture your code sends is. The iOS and Android SDKs, built the same way, are in development.
Privacy you set, per organization
Four settings decide what ReceViz keeps and where document text may go. An application can only make its organization's settings stricter. They are controls, not a certificate: the documentation says exactly what each one does, so you can judge how they fit your obligations.
- Result retention
- Results are kept for 30 days or 24 hours, or handed over once and not kept. The uploaded file is never kept once it has been read.
- Evidence
- Return the printed text and its position with each value, the position only, or the values alone.
- Language models
- Allow document text to reach a language model, or keep it away. With them off, VERIFIED mode is refused rather than quietly weakened.
- Quality telemetry
- Record nothing; record per-field agreement that never holds a value or a line of text; or also keep the text of a misread, only when a request explicitly shares it.
Try it on your own documents
ReceViz reviews every request for access. Once your organization is approved, you create keys and document types in the console and try them in its Playground before writing any code.