Receipt reading

Receipt reading turns a photo or a PDF of a receipt into a prefilled expense: amount, date and merchant. It is off by default, optional, and entirely local. What members see is on Scanning receipts.

#Local by design

The receipt is read on your server by Tesseract, an open-source OCR engine, then parsed into a total, a date, a merchant and line items. The file lives in memory for the length of the request and is never written to disk or the database; only the reading goes back to the browser. Nothing is sent to any outside service.

Reading happens while the person waits, with a two-minute limit per file. Files are capped at 10 MB, in JPEG, PNG, WebP or PDF.

#Turning it on

Admin, Receipts shows what this server has:

• whether Tesseract is installed, and which language packs it has;
• whether PDFs can be read, which needs Poppler's pdftotext and pdftoppm next to Tesseract. The Docker image ships both.

Pick the languages your receipts are printed in and save. Reading in several languages at once works, but each extra one makes it a little slower and can make it a little less accurate, so pick only what you need.

Try a receipt reads a file right there and shows everything: merchant, date, total, the line items it found and the raw text Tesseract recognized. Use it to check the languages before members rely on them.

Turn off receipt reading removes the Scan a receipt button from the expense sheet.

#Languages

The published image ships English, French, German, Spanish and Italian (eng fra deu spa ita). To ship others, build the image with the OCR_LANGUAGES build argument, listing Alpine tesseract-ocr-data-* package suffixes:

docker buildx build --build-arg OCR_LANGUAGES="eng fra nld por" \
  -t opencount:latest .

Outside Docker, install Tesseract and its language data yourself (apt install tesseract-ocr or brew install tesseract tesseract-lang), and Poppler for PDFs (apt install poppler-utils or brew install poppler).

#PDFs

A PDF is read from its first page. When the page carries text, like a receipt emailed by a shop, that text is used directly. When it doesn't, like a scan, the page is rendered at 300 dpi and read by Tesseract. Both go through pipes; nothing touches the disk.

#How the parsing works

• Total: the largest amount on a line that says total in one of the languages it knows (total, montant, summe, betrag, importe, à payer, amount due and more), skipping subtotals. With no such line, the largest amount on the receipt.
• Line items: lines with words and an amount, minus tax, change, tips and card lines; at most 50.
• Date: ISO dates first, then day-first local dates.
• Merchant: the first line with at least three letters and no amount.
• Currency: from a symbol or code like €, £, CHF or zł. It formats the total in the admin test but is not applied to an expense, whose currency is the group's.