Extract tables from PDF
Export document tables to spreadsheets.
Finds the tables in a PDF, lets you correct them beside the page they come from, and exports them as Excel, CSV, JSON or Markdown.
How to use it
- Choose a PDF with selectable text and, optionally, the pages to read. For scanned PDFs, run OCR PDF first.
- Select Find tables. Each table appears with its page; a table that runs onto the next page is joined, and Page 2 shows the rest.
- Correct cells by typing in them. Add or remove rows and columns, choose whether the first row is the header, and use Join wrapped lines or Clean up with AI when superPDF says a table looks irregular. Undo reverses each change.
- Download a table as CSV, or every table you leave ticked as one Excel workbook (a sheet per table), JSON or Markdown.
Options
| Option | Values | Default |
|---|---|---|
| Pages | Page numbers or ranges, for example 1, 3-5 | All pages |
How tables are found
Finding tables uses no AI: it is the same reconstruction as PDF to Excel, run in your browser. Where the layout can’t be split cleanly, such as rows that wrapped onto two lines or columns that ran together, join wrapped lines yourself or ask the AI to clean the table up.
Clean up with AI
The AI is given the lines of text the table came from, with the gaps superPDF found marked, and the table as superPDF read it, and is asked to join wrapped cells, split cells that hold two columns, and change nothing else. It uses the AI you chose in Chat with PDF: your own API key (the table’s text goes from your browser to your provider) or a model on this device. Every cell of the result is then checked against the page text: a cell whose words or number aren’t on the page is cleared and marked, and words of the original table that the result leaves out are listed. With a model on this device, tables up to about 3,500 characters can be cleaned up; with your key, up to about 40,000.
Output
- Excel: numbers are stored as numbers, as in PDF to Excel; a header row is bold and stays in view.
- CSV: one table per file. Text that a spreadsheet would treat as a formula (starting with =, +, - or @) gets a leading apostrophe, so a document can’t plant a formula in your spreadsheet.
- JSON: each table’s pages, its header row (or
null) and its rows, with every cell as text exactly as shown. - Markdown: a table per heading, ready to paste into notes or documentation.
Limits
superPDF sets no file size or page limit. Grid lines aren’t read, so columns need visible space between them. A cell that spans several columns goes in the first of them. The AI cleanup can still misplace a value into the wrong row or column: it can only use text from the page, but check the result against the page beside it.
Scanned pages: when superPDF offers high-accuracy reading, the tables it reads from the page pictures are added as tables of their own.
High-accuracy reading
When a PDF has scanned pages, tables or math, superPDF can offer high-accuracy reading beside the quick result. It reads each page as a picture with a small AI model made for documents (LightOnOCR-3 by LightOn), so it gets the words, tables and math of scans right far more often than the quick reader. It is offered on computers whose browser can run AI on the device (the latest Chrome, Edge or Safari) with 8 GB of memory or more; phones and tablets keep the quick reader. You see the quick result first, then each page is replaced as it is read, with the time left and a Stop button. Once you have chosen it, superPDF uses it by itself in this browser where it helps. It runs on this device: nothing leaves it, and it works offline once downloaded. Download: 871 MB, once, only when you choose it; delete it any time in On-device AI models. Speed: about 10 to 30 seconds a page on a recent computer (the offer shows the time for your pages); up to 20 pages are read at a time unless you choose more or pick pages. It reads English and other European languages best; check results in other scripts, and like any OCR it can misread faded or handwritten text. Running page headers, footers and page numbers are left out of the text. If anything goes wrong, the quick result stays, with a short note.
Privacy
Finding and editing tables runs in your browser with no AI. Clean up with AI uses the AI you chose: with a model on this device, nothing leaves it; with your own API key, the text of that table goes from your browser straight to your provider.