How PDF2Sheet Extracts Tables — And Why Nothing Is Uploaded

Step 1 — read the text, and its position

The first thing that happens is that PDF.js, the open-source PDF engine behind Firefox's built-in viewer, parses the file. Importing it from this site's own /vendor folder is the only third-party code that runs.

PDF.js returns a list of text runs. Each run carries a string and a transform matrix — in practice, an x and y coordinate and a font size. That is the raw material:

str: "POS PURCHASE - WHOLE FOODS MKT" x: 130.0 y: 214.5 width: 132.4 height: 8.0

Note what is not in that list: rows, columns, borders, or any idea that these runs belong together. A PDF is a page description, not a data format. Everything after this point is reconstruction.

Step 2 — group runs into rows

Runs whose baselines sit within about a third of the font size of each other are treated as one line. The tolerance scales with the text: a 20 pt headline and an 8 pt caption still count as the same line if their baselines are two points apart, but two lines of 8 pt body text twelve points apart do not.

Inside a line, runs separated by less than roughly four-tenths of an em are joined — that is just a space in the same cell. Wider gaps stay as separate cells.

Step 3 — find the columns

This is where most converters go wrong, so it is worth being specific about what this one does. It clusters the left edge of every cell into column anchors, and assigns each cell to its nearest anchor.

Three consequences follow, and they are the reason this approach was chosen:

• A wide cell — "Closing balance", a long merchant description — starts in its column even though its middle extends past the next one. Centre-based assignment gets this wrong; left edges do not.
• A header printed on the same baseline as body text (a letterhead that says ACME BANK   Statement period) cannot pollute the result, because each row votes for anchors independently.
• A sparse column survives. If only three rows have a value in the credit column, three cells starting at the same x are still strong evidence of a column.

Anchors need a minimum amount of support to exist, so a single stray label floating in whitespace does not create a column of its own.

Step 4 — decide what is actually the table

A page has furniture: bank name, address block, page numbers, legal text. The converter keeps the densest run of numeric rows — real table bodies contain amounts and quantities, letterheads do not — and pulls in the header line directly above it. Everything else is dropped when Only export the table is on.

Step 5 — clean up the values

Finally each cell is examined. Amounts become numbers: brackets mean negative, a trailing minus means negative, both 1,234.56 and 1.234,56 are understood, and ISO currency codes in front of a value are stripped. Identifiers that merely contain digits — account numbers, invoice numbers, ZIP codes with a leading zero — are deliberately kept as text so their leading zeros survive.

How to verify that nothing is uploaded

You do not need to trust this page. Two minutes with your browser's developer tools settles it.

1. Open the Network panel

   Press F12 (or Cmd+Opt+I on a Mac), choose the Network tab, then reload the page so you start from a clean list.

2. Clear the list, then convert a file

   Click the clear button in the panel, then drop a PDF into the converter above. Watch the list as it runs.

3. Read what appears

   You will see this site's own files: the HTML, the stylesheet, app.js, engine.js, the PDF.js worker, the spreadsheet writer. No request carries your document. There is no POST, no multipart/form-data, no WebSocket.

The same check works offline: load the page, then disable your network adapter, and convert a file. It still works, because the conversion never needed a network in the first place.

Known limits, stated plainly

• Scans and photos. No text layer, no output. Detected and reported rather than silently returning an empty sheet.
• Columns with no whitespace between them. If two fields touch, there is no geometric signal to split on. The More columns setting can help when the gap is small but non-zero.
• Tables drawn as images inside a text PDF. A chart or a screenshot of a table is a picture, not text.
• Multi-line cells. A wrapped description is merged back into its row when the continuation has no date of its own. If the bank repeats the date on every line, the merge cannot fire.
• Dates are kept as printed. A statement that says 02 Sep without a year stays text; inventing a year would be worse than leaving it.

What the code is

The extraction engine is a single dependency-free JavaScript file of roughly 600 lines: row clustering, column anchoring, table detection, number parsing. PDF.js and SheetJS are used unmodified. There is no analytics script on tool pages, no tracking pixel, no cookie set by the converter.