Extract the line-item table from a scanned invoice

What this tool does, exactly

An invoice is a table with a story: lines, a subtotal, tax, a total due. This page reads the table out of a scan or a photo with the Deep reader and writes it as an .xlsx. Then it does the one check a spreadsheet can do that a reader cannot: it adds the lines up and says whether the subtotal and the total agree with them. A total that does not add up is the fastest sign that a digit was misread.

What the totals check is

Arithmetic on the rows the reader produced, nothing more. For every row whose label says total — Subtotal, Total, Total due, Amount due — each column's figure is compared with the sum of the lines above it, back to the previous total row. It is also compared with that previous total carried forward with the lines between, so Subtotal + Shipping + VAT = Total due; whichever adds up is the reading reported. A match within rounding — half a cent a line — is reported as adding up; a mismatch names the row and both figures. It says whether the reading is consistent with itself. It is not accounting, and it does not know whether the invoice is right.

Currency and numbers

A cell that is a bare figure — 220.80, or 1,347.72 with thousands commas — becomes a number in the sheet. A cell with a currency mark, like € 220.80 or 220.80 EUR, stays text, because a number stripped of its currency is a different fact; the arithmetic reads past the mark all the same. A European decimal such as 1,5 stays text rather than becoming fifteen.

Crop first, and why

Not for speed but for structure. The reader spends about the same time on a picture whatever is in it, because the picture is scaled to one size first: measured here, a whole A4 invoice took 38 seconds and its table alone 39. On the whole page the reader turned the totals block sideways — labels in one row, figures in the next — and invented an empty sixth column. On the cropped table every row came back right and the totals added up. A cropped table is read at higher resolution for the same seconds.

What it refuses

A phone or tablet; a device with under 8 GB of memory; a browser without WebGPU, the graphics-chip interface; a document of more than one page. And a picture the reader does not answer with rows and columns — a receipt printed as a single column of text is read as text, not invented as a table. Each refusal names itself.

What it costs

The Deep reader, about 660 MB from Hugging Face's servers once, kept in this browser afterwards; then fifteen to forty seconds a table on a recent laptop's graphics chip. Free during the preview, behind the Pro door: no card, no account.