Receipt data extraction

Where receipt extraction fits

Receipts are the small end of finance paperwork: expense claims, card reconciliation, and bookkeeping for firms that collect shoeboxes of them from clients. In each case someone is matching a slip of paper to a card line and a tax treatment, and the amounts are small enough that keying them feels absurd and large enough in aggregate that it can't be skipped.

The evidence standard is modest but real. For US business expenses, the IRS says documentary evidence will ordinarily be adequate if it shows the amount, date, place and essential character of the expense. Receipt Date, Supplier Name, Supplier Address, Total Amount and the item lines map onto that list. Whether a claim is deductible is your accountant's call, not ours.

Receipt fields docRefinery extracts

There are 35 definitions, including the item table. Till details such as Terminal ID and Authorization Code help tie a receipt to its card transaction.

Example receipt JSON

A trimmed excerpt with invented values. The handwritten tip scored under its threshold, so this run waits for a person. For the complete response, see Extract a document.

{
  "document_type": "receipt",
  "result": {
    "fields": [
      { "alias": "Supplier Name", "text": "Sample Street Cafe", "field_confidence": 0.93,
        "location": { "page": 1, "bbox": [0.24, 0.05, 0.76, 0.08] } },
      { "alias": "Total Amount", "text": "18.40", "field_confidence": 0.96,
        "location": { "page": 1, "bbox": [0.64, 0.71, 0.86, 0.74] } },
      { "alias": "Tip Amount", "text": "3.00", "field_confidence": 0.62,
        "location": { "page": 1, "bbox": [0.60, 0.78, 0.80, 0.82] } },
      { "alias": "Supplier Tax ID", "text": null, "location": null },
      { "alias": "Line Item", "table": [[
        { "alias": "Description", "text": "FLAT WHT LGE", "field_confidence": 0.88,
          "location": { "page": 1, "bbox": [0.10, 0.33, 0.52, 0.35] } }
      ]] }
    ],
    "extraction_approval": "human-approval",
    "low_confidence_fields": [
      { "alias": "Tip Amount", "text": "3.00", "confidence": 0.62,
        "threshold": 0.8, "reason": "low_confidence" }
    ]
  }
}

What makes receipts hard to read

• Thermal paper. Its coating stays active after printing, so heat, light and friction can fade a receipt or darken it. A claim filed weeks late may be reading a ghost.
• The photo itself. Creases, shadows, curled edges, a thumb over the corner, and a long narrow strip shot at an angle.
• Till abbreviations. Item names are cut to fit the printer, so FLAT WHT LGE is what the page says.
• Handwriting on print. Tips and revised totals are often written in by hand beside the printed figure, so two candidate totals sit on one slip.
• Several numbers that look like the total. Subtotal, total, amount tendered and change usually stack together.
• Travel. Foreign currency, local tax names and unfamiliar scripts turn up in the same expense claim.

How low-confidence receipt values are handled

Faded print usually fails quietly: the digit is still there, just not certain. So each value is scored against its own threshold, and a total scored under the bar goes to the review queue with a box around the patch of the photo it came from. An expenses clerk confirms or corrects it in seconds, with no hunting along the strip. A tax ID that was never printed returns null, not an invented number. You might accept looser reads on Cashier Name and hold Total Amount tight; thresholds are set field by field.

What docRefinery adds for expenses and bookkeeping

A single receipt is worth little, so what matters is how little handling it needs once it has been read.

• Expense tools get their own labels. Set Return as on Supplier Name to merchant and on Total Amount to amount_gross, and the API and CSV headers carry those names.
• Receipts and invoices share one output format, so a bookkeeper's import handles both without a second parser.
• Items come back as rows, ready to split across categories in your own system.
• A client's stack scanned as one long PDF is segmented into the documents inside it.
• Slips from a trip abroad are covered by 70+ languages and scripts.
• A clerk's correction on a regular merchant becomes a suggested fix, tested on that layout's own receipts before anyone applies it.
• The Chrome extension, in early access and not in the Chrome Web Store yet, runs a receipt and auto-fills a web form such as an expenses portal.
• Starter is free up to 100 pages a month with no card, enough to try a real month of slips.

Where receipts fit docRefinery less well

Layout memory is our main advantage, and receipts test it. Regular chains print from a template, so an account's usual merchants are learned once and remembered. A traveller's one-off stops are all first sightings. That long tail still gets your field names, item rows, scoring, review and the same output format, with less of the benefit that builds as layouts repeat.