PDF to CSV: pull tables out of a text PDF
PDF to CSV works when the PDF has a real text layer.
Updated
How it works
Add a text PDF
Open the tool (it may ask you to sign in) and drop in one PDF where you can select the text.
Choose CSV or Markdown
Run Extract tables. The cost is shown before you send. Each table comes back as its own block with its page number.
Download and check
Press Download .csv, open it in your spreadsheet and compare the totals with the PDF.
How do I get a table out of a PDF?
Add the file, pick Extract tables and choose CSV or Markdown. The model reads the text layer, finds the tables and writes each one as a separate block, labeled with the page it came from. A PDF with six tables gives you six blocks, not one long pile of rows.
CSV is what you want for a spreadsheet. Markdown is handy if you're pasting into notes or a document. Each CSV block has a "Download .csv" button.
For a longer walkthrough with a worked example, see PDF to CSV with AI.
What to ask
Extract every table in this document as CSV. Put the page number above each one.
Use the tool aboveExtract only the table that lists monthly charges, as CSV, with the header row first.
Use the tool aboveUse plain numbers without currency symbols or thousands separators, and keep the dates as they appear.
Use the tool aboveIf any cell was merged or split over several lines, list the row and tell me how you read it.
Use the tool aboveHow accurate is it?
Good on tidy tables, shaky on messy ones. Merged cells and cells that wrap over several lines are where it goes wrong: a value can slip into the wrong column or two rows can fuse into one.
So check the totals yourself. Add up a column in your spreadsheet and compare it with the total printed in the PDF. If they differ, find the row. It takes a minute and saves you from a wrong report.
If a table comes back crooked, ask again for just that table and say what's off ("the third column is missing").
What it will not do
- It only reads text PDFs. A scan or a photo of a table gives "No text found", and there's no OCR.
- It won't promise perfect layouts for merged or multi-line cells.
- It can't open a password-protected file until you remove the password elsewhere.
- It won't build formulas or charts. You get rows of text.
What does it cost, and what leaves my device?
Extract tables runs on Claude Haiku 4.5 at 5 Gems per message. New accounts start with 30 Gems. See pricing for packs and plans.
The PDF is uploaded to the model provider you picked, and sent again with each message. If the tables sit on a few pages of a big file, cut those pages out first with the split tool. It runs on your device.
Questions about PDF to CSV
Can I convert a scanned PDF table to CSV?
No. The tool reads the text layer only, and there's no OCR. A scan or photo returns "No text found". If you can select and copy the numbers in your viewer, the file should work.
Why are some columns shifted in my CSV?
Merged cells and cells with several lines of text confuse the reading. Ask again for that one table, describe the problem, and compare the result with the page before you use it.
Can I get the table as Markdown instead?
Yes. When you run Extract tables you can choose Markdown instead of CSV. The output is still one block per table with its page number. The Download .csv button is for the CSV option.
Does it handle a PDF with many tables?
Yes, each table gets its own block with its page. Very long files can be cut or refused, since a PDF counts as roughly 6,000 input tokens. Pick a page range if you only need part of it.
Should I trust the numbers it extracts?
Treat them as a first draft. Models can drop or misplace a digit. Add up a column and compare it with the total in the PDF before the data goes into anything that matters.