How do I get the tables out of a PDF as rows and columns?

A PDF has no tables in it. It has text placed at coordinates, and a table is something a reader infers from how that text lines up. Which is why copy-pasting a PDF table gives you a run of words: the structure was never stored.

Rebuilding it means grouping text by its position — and doing it from the alignment rather than from the drawn lines, because plenty of tables have no borders at all. What you should get back is an array of rows, each an array of cells, with the page it came from.

A real run

{
  "pdfUrls": ["https://www.irs.gov/pub/irs-pdf/i1040tt.pdf"],
  "detectTables": true,
  "extractKeyFields": false,
  "outputFormats": ["json", "csv"],
  "maxPages": 10
}

{
  "url": "https://www.irs.gov/pub/irs-pdf/i1040tt.pdf",
  "ok": true,
  "pageCount": 28,
  "pagesCharged": 10,
  "tableCount": 34,
  "tablesCsvUrl": "https://api.apify.com/v2/key-value-stores/.../records/001-i1040tt.csv?signature=...",
  "jsonUrl": "https://api.apify.com/v2/key-value-stores/.../records/001-i1040tt.json?signature=...",
  "metadata": {
    "title": "2025 Publication 1040",
    "author": "W:CAR:MP:FP",
    "subject": "TAX AND EARNED INCOME CREDIT TABLES",
    "keywords": null,
    "creator": "AH XSL Formatter V6.6 MR4 for Linux64 : 6.6.6.37929 (2019/03/11 12:35JST)",
    "producer": "Antenna House PDF Output Library 6.6.1437 (Linux64); modified using iText 2.1.7 by 1T3XT",
    "createdAt": "2025-09-17T17:52:35.000Z",
    "modifiedAt": "2026-01-15T07:41:41.000Z",
    "pageCount": 28,
    "pdfVersion": "1.7",
    "language": "EN",
    "encrypted": false,
    "hasAcroForm": false,
    "hasSignatures": false,
    "linearized": true
  }
}

Each table comes back with its shape and its position, so you can tell a real table from a layout artefact before you try to use it:

{
  "page": 9,
  "rows": [
    ["Your tax is—", "Your tax is—", "Your tax is—"],
    ["57,000", "60,000", "63,000"]
  ],
  "header": ["Your tax is—", "Your tax is—", "Your tax is—"],
  "rowCount": 2,
  "columnCount": 3,
  "bbox": { "x0": 54.2, "y0": 109.1, "x1": 532.4, "y1": 134.5 }
}

That is a small one, picked because it fits on this page. The tax tables themselves come back as 18-column, 21-row grids — ["57,000", "57,050", "7,460", "6,366", …] — and the same ten pages exported to CSV are 636 lines with a # table n — page n comment before each one.

What else is in a PDF worth taking

• Text in reading order, including multi-column pages read column by column rather than straight across — plus Markdown with the headings kept, if you are chunking for retrieval.
• The outline: the bookmark tree, each entry with its level and the page it points at. The cheapest way to see how a long document is organised before deciding which pages to parse in full — one page is enough to read it.
• Form fields of a fillable PDF: name, type and current value, empty ones included, so you get the structure of the form as well as what somebody filled in.
• Link annotations: the real clickable links with their target and position, not URLs guessed out of the text, so a link whose visible label differs from its destination is still reported correctly.
• Key fields from an invoice — number, dates, totals with currency, VAT ID, IBAN. IBANs are checksum-verified and an ambiguous date is flagged rather than guessed.

What this does not do

• No OCR. A scanned page is an image, and this reads the text layer. Scanned pages are flagged as such — and not charged — but their words do not come back. If your PDFs are scans, you want an OCR tool, not this one.
• Table detection is inference, not truth. On a page where the layout itself is built out of aligned text, you will get a “table” that is really a layout. The row above is an example: it is a column header repeated three times. rowCount, columnCount and bbox are there so your code can filter those out.

What it costs

$0.002 per page extracted, capped at 40 pages per document. The run above touched 10 pages of a 28-page file and was charged for 10 — firstPage and maxPages mean you pay for the range you need and not for the document. Reading only the metadata or the outline costs a single page. Scanned pages and files that will not download are free.