Skip to content

Every field, with its source.

Structured data from invoices, statements and forms. Each value is read twice, checked against the page by a judge, and returned with its page, quote and box.

halden-rowe-INV-2026-0142.pdfSynthetic invoice

FieldsInvoice

2 to check

  • SupplierHalden & Rowe Stationers Ltd97
  • Invoice no.INV-2026-014299
  • Issue date2026-09-0362Doubt
  • PO numberPO-5581784Check
  • Total391.92 GBP98

Found03/09/2026onp.1near“Date”

03/09 can be 3 Sep or 9 Mar.

PO numberPartly under the “Received” stamp.

How a value earns its confidence

One field from the invoice above, followed through the pipeline.

  1. 1. Drop the files

    Up to 30 at once. No account and no schema: the document type is recognized and the fields are proposed from the page.

    • halden-rowe-INV-2026-0142.pdf1 page
    • receipt-taxi-0911.jpgphoto
    • lease-unit-7.docxWord
  2. 2. Read twice

    Two independent passes, with different prompts and models, read every field. Agreement raises confidence; disagreement marks the field.

    Pass Aissue_date2026-09-03
    Pass Bissue_date2026-03-09

    Readings differ

  3. 3. Checked against the page

    A separate judge confirms each value against the quoted text on its page. Totals, VAT, NIF and IBAN are checked by arithmetic.

    Judge

    Quote “03/09/2026” is on p.1. Both readings fit it.

    Σ lines = subtotal326.60 = 326.60
  4. 4. Source and confidence

    Every value comes back with its page, quote, box and a confidence from 0 to 100. Flags mark the few to check.

    Issue date2026-09-0362Doubt

    Day and month order is ambiguous.

A confidence you can act on

The point is not that the model is good. The point is that you can see, per value, how sure the system is and why.

Evidence on the page
Each value carries the page number, the exact quote and a box on the page. Select a field and the page shows where it came from.
A number with a method
Confidence comes from agreement between the readings, the judge's verdict and the arithmetic checks. Not from the model grading itself. 90+ no flag, 70 to 89 Check, below 70 Doubt.
Nothing hidden
A value the document does not contain comes back empty and flagged, never guessed. Edits are marked as reviewed and carried into every export.
The JSON export, per fieldexport.json

"fields": {

"supplier_name": "Halden & Rowe Stationers Ltd"

97

"invoice_number": "INV-2026-0142"

99

"issue_date": "2026-09-03"

62Doubt

"po_number": "PO-55817"

84Check

"total": 391.92

98

}

{

"key": "issue_date"

"type": "date"

"value": "2026-09-03"

"raw": "03/09/2026"

"confidence": 0.62

"status": "needs_review"

"reason": "Day and month order is ambiguous"

"alternatives": [{ "value": "2026-03-09" }]

"page": 1

"bbox": [0.742, 0.271, 0.931, 0.302]

"reviewed": false

}

Documents it knows

Built-in field sets for common documents. For anything else, the fields are proposed from the page.

Exports that keep the evidence

Every export carries confidence, status, page, box and a reviewed mark per value. Per document or per batch.

extraction-2026-10-23.xlsx

documenttotaltotal__confidencetotal__status
INV-2026-0142.pdf391.920.98verified
INV-2026-0155.pdf88.560.99verified
receipt-taxi.jpg41.200.74needs_review
bill-sept.pdf61.030.96verified
fieldsline_itemsreview

Excel

One workbook: a fields sheet with a row per document, a sheet per table (line items, transactions), and a review sheet listing every value with its status and reason.

fields.csv

document,total,total__confidence,total__status,total__page,total__bbox

INV-2026-0142.pdf,391.92,0.98,verified,1,"0.71,0.75,0.93,0.78"

INV-2026-0155.pdf,88.56,0.99,verified,1,"0.70,0.81,0.93,0.83"

receipt-taxi.jpg,41.20,0.74,needs_review,1,"0.52,0.77,0.80,0.80"

+ table-line_items.csv

CSV

fields.csv with a row per document, plus one CSV per table, in a ZIP. UTF-8 with BOM, so accents open correctly in Excel.

export.json

{

"key": "total"

"type": "money"

"value": 391.92

"raw": "£391.92"

"confidence": 0.98

"status": "verified"

"page": 1

"bbox": [0.71, 0.75, 0.93, 0.78]

"reviewed": false

JSON

Typed values with raw, confidence, status, page and bbox for every field and every table cell, plus the checks that ran.

  • Templates

    Save a field list and run a batch with it: same columns for every file, one consolidated table.

  • Share links

    Send results to someone without an account, with an expiry date.

Questions

Do I need an account?

No. You can drop up to 3 files a day without one, and the first 3 pages of each file are read. A free account raises that to 10 files a day. If you sign up later, the files you already uploaded move into your account.

Which files can I upload?

PDF (with a text layer or scanned), JPG, PNG, WEBP and GIF images, Word (DOCX), Excel (XLSX), CSV and plain text. Up to 30 files at once, 50 MB each.

Do I have to tell it which fields to extract?

No. It recognizes common documents (invoices, receipts, bank statements, utility bills, payslips, contracts, forms) and proposes the fields for anything else. If you want the same columns across many files, save the field list as a template and run the batch with it.

What does the confidence number mean?

It is computed from three things: whether two independent readings agree, whether a separate judge confirms the value against the page, and whether arithmetic checks pass. It is not the model rating itself. 90 and above gets no flag, 70 to 89 gets a Check flag, below 70 a Doubt flag.

What happens when a value is wrong?

You fix it in place on the review screen, next to its source. The value is then marked as reviewed by a person, and that mark goes into every export.

Where are my files stored?

In the EU. Files are kept in your account until you delete them. Documents are sent to the model providers only to process your request and are not used to train models. The privacy page has the details.

Is there an API?

Not yet. Today you upload in the browser and download JSON, CSV or Excel. The JSON export has the same shape an API would return: value, confidence, status, page and box per field.

Run it on a document you know

Pick a file where you already know the right answers. Compare, field by field, with the source on screen.