doc2data: invoices and receipts to checked data

Invoices and receipts arrive as PDFs or scans and are retyped by hand into spreadsheets.

Extract the fields without guessing: every uncertain value has to land on a review list, not in the final spreadsheet.

What it extracts

Vendor, VAT number, invoice number, date, currency, net amount, VAT, total, and every line item.

How it checks

• Net + VAT must equal the total; line items must add up
• Dates must be plausible; VAT numbers must have a valid format
• Anything that fails a check is flagged for review, never guessed

How it runs

• Offline, on your own computer, with no cloud service
• Text PDFs are read directly; scans go through OCR (optical character recognition) on your machine, with no cloud service
• An AI fallback for unusual layouts is optional and off by default. It asks only for fields the rules missed, and those fields are always flagged. The AI fallback is off by default. If you switch on the cloud option, it sends up to the first 6,000 characters of the extracted text to an OpenAI-compatible provider you choose, using your own API key. The optional Ollama fallback runs on the same machine (localhost); nothing leaves the computer. Both are built and unit-tested, not yet tested against a live model. The demo app offers only the local option; the cloud option exists only in the command-line tool.

Measured on 2 October 2026 in our Docker build. All documents are synthetic: invented companies, addresses and VAT numbers.

• The invoice with a deliberate arithmetic error was flagged in every set.
• The automated test suite passes 21 of 21 tests.

Full test evidence: field counts, review flags and limits for all five sets

Synthetic documents. Real accuracy must be measured on your samples. The synthetic invoices use 2 layouts plus a receipt format, and the rules were written knowing them. Real supplier documents will score lower; that is why every project starts by measuring on your own documents.

Full test evidence

Python, pdfplumber and pypdfium2 (PDF text), RapidOCR on ONNX Runtime (offline OCR), pandas, openpyxl, Streamlit; synthetic test PDFs generated with ReportLab; tests with pytest; packaged with Docker.

• Several VAT rates on one invoice: not handled
• Line-item tables across several pages: not handled
• Credit notes: not handled
• Line descriptions that wrap onto two lines: not handled reliably
• VAT numbers are checked for format only, not by checksum or against the EU's VIES register
• Single-character OCR errors in names and VAT numbers on poor scans cannot be caught by arithmetic checks
• No real-world accuracy has been measured yet