Invoice Data Extractor

How to use Invoice Data Extractor

1. Choose the invoice or receipt — a PDF, or a photo of a till receipt.
2. Press Extract the fields. A PDF with real text is read directly; a scan or a photo is read as a picture first, which takes longer.
3. Check the numbers against the document, then take them as a table, CSV or JSON.

Questions

Does my invoice get uploaded?

The text of it does. Finding fields in an unlabelled document needs a language model, and that runs on a provider, not in your browser — so this is one of the few tools here that sends anything. The file itself is not stored, and the text is not kept after the answer comes back. If your invoice is a structured XRechnung or ZUGFeRD, use the E-Invoice Viewer instead: those are already labelled, so it reads them in your browser with no upload at all.

Why is a field empty?

Because it was not found in the document. The model is instructed to return nothing rather than guess, and that instruction is the point: an invented VAT number that looks right is worse than a blank one you can see and fill in.

What does “the numbers do not agree” mean?

After the fields come back, the arithmetic is re-done here — net plus VAT against the gross, the VAT rate against the VAT amount, and the line items against the net total. When any of those disagree by more than a cent, the page says so. It usually means the document carries more than one VAT rate, or that a line was not read.

My invoice is in German. Does that work?

Yes — pick the language before you press the button. It only matters for a scan or a photograph, because a PDF carrying real text is read directly and never goes near the character recogniser. On a photographed German invoice it is the difference between Müller Großhandel and "Miller GroBhandel": read as English, every umlaut and every ß is lost. German, French, Spanish, Portuguese, Arabic, Hindi, Bengali, Russian, Chinese, Japanese and Korean are all there.

Can it read a photo of a receipt?

Yes, within reason. The picture is read with OCR first and the text handed on, so the characters have to be legible to a person before they are legible to anything else. A sharp, straight, well-lit photo works; a crumpled receipt at an angle in low light does not.

Can I run a whole folder through it?

Yes, up to 50 at a time. Because each document is a charged run, it counts them first and tells you how many are free today and how many come out of your balance — nothing is charged until you confirm. They are read one at a time, and a document that cannot be read is listed with its reason rather than quietly skipped, so 50 files always produce 50 rows.

What do I get back from a batch?

A spreadsheet: one row per document, with the line items counted rather than listed, which is what keeps the file sortable. Download it as a single CSV. A row marked ⚠ is a document whose own totals did not re-add — worth opening on its own, where you also get the invoice displayed beside its fields.