PDF を CSV に変換

このツールでできること

A PDF has no idea what a table is. There are no cells and no rows in the file, only text placed at coordinates that happen to line up, so recovering a table means finding the columns again. This tool clusters the horizontal positions of every text run, looks for a stretch of lines whose runs all fall into the same bands, and writes what it finds as CSV. Because it is inference rather than reading, every table comes with a confidence score, and the threshold is yours to set.

The usual case is numbers you would otherwise retype: a bank statement that only arrives as a PDF, invoice lines that have to be reconciled, a price list from a supplier, or an export from a system that offers PDF and nothing else.

仕組み

1. Drop the PDF onto this page. If it is a scan with no text layer, you are sent to OCR PDF rather than given an empty spreadsheet.
2. Pick the delimiter. Comma is standard; semicolon is what Excel expects in most of Europe; tab suits pasting straight into a sheet.
3. Set the confidence threshold. 55 per cent is the default; lower it to catch loosely aligned tables, raise it if prose is being mistaken for one.
4. Choose one file per table, or a single file with the tables one after another.
5. Press Extract tables. Each output shows its page and its confidence, so a doubtful table is visible as doubtful before you use it.

The detail that separates a table detector that works from one that does not is right alignment. Number columns are set flush right, so their runs share an end position rather than a start position, and a detector that clusters only on where text begins misses every financial table - which is most of what anybody converts. Both edges are clustered here.

Nothing is returned below the confidence threshold, and that is a deliberate refusal rather than a gap. A spreadsheet full of misaligned fragments looks like a result and is worse than no result, because someone will use it.

The threshold is a real dial. At 30 per cent almost any run of aligned lines is offered, which is useful for a loosely typeset table you already know is there. At 90 per cent only tables with clean, consistent columns survive, which is what you want when converting a hundred pages unattended. The default of 55 is where a typical report's tables come through and its paragraphs do not.

Detection and writing both happen in your browser. The numbers in a bank statement are about as sensitive as a document gets, and none of them are transmitted: the PDF is read and the CSV written without any part of either reaching a network.