PDF Tables to JSON & CSV

Give it links to PDFs. It returns every table it finds as clean, typed rows, ready to export as CSV, Excel or JSON.

• CSV-ready by default. Each table row is one dataset item and the table's columns are its fields, so an export is a flat table. _sourceUrl, _table, _page and _row say where each row came from. To get one item per table instead, set outputMode to tables.
• Numbers come back as numbers. $1,234.50 becomes 1234.5, (45.00) becomes -45.0, 12% becomes 0.12 and 40 stays 40. Identifiers stay text: 00123 and long digit strings are not turned into numbers.
• A table that continues on the next page comes back as one table when its header repeats there, which is how most generated reports print it.
• Two tables on one page come back separately. Each has its page, its position and its size.
• A confidence score per table. For ruled tables it is the share of cells on which two independent readings agree. For tables with only header rules it is how consistent the row widths are. Correctly read tables score 0.9 or more. A misread one, such as values in the wrong columns, scores low.
• Pay only for PDFs where a table is found. PDFs with no table, refused links and failed downloads are free. Each still leaves a record (_status) saying why.

Input

{ "startUrls": [{ "url": "https://example.com/price-list.pdf" }], "outputMode": "rows", "maxPages": 100 }

Public URLs only (file upload is not in this version). Up to 500 PDFs per run, 1-300 pages each, 50 MB per file.

Three optional fields are under "Help improve this tool" and are described in What this tool collects: useCase (what kind of documents these are), shareAnonymousStats (on by default, uncheck to turn it off) and shareFailingLinks (off by default).

Output (rows mode, one item per row)

{ "_sourceUrl": "https://example.com/price-list.pdf", "_status": "ok", "_table": 0, "_page": 1, "_row": 0,
  "_confidence": 1.0, "SKU": "SK-1000", "Item": "Bolt 12 mm", "Qty": 40, "Unit price": 18.5, "Discount": 0.1 }

Pricing

$0.01 per PDF where at least one table is found, charged after all its rows are saved. A PDF with no table, a link that is refused or a download that fails costs nothing. Apify also charges a fixed $0.00005 each time the tool starts, however many PDFs that run reads.

What this tool collects

At the end of each run the tool sends the developer one anonymous record, unless you turn it off. It holds:

• the date of the run (UTC day only), and the version number of the record format;
• the tool's version, your outputMode, and a size band for maxPages (for example "51-100");
• your answer to "What are these PDFs?", if you gave one;
• how many PDFs ended each way: tables found, tables found with some pages skipped, no table, failed, link refused, stopped at the spending limit;
• the number of pages read and tables found, and how many tables scored below 0.5, from 0.5 to 0.9, and 0.9 or more;
• a band for how long the run took (for example "10-60s");
• which program made the PDFs (Word, Excel, LibreOffice, LaTeX, InDesign, Acrobat Distiller, Chrome, and so on, or "other" or "unknown"), counted per outcome. The tool reads the PDF's own producer field and keeps only the name of the group it falls in, never the text of the field.

It never holds your links, their domains, file names, anything written inside a PDF, your Apify account or user id, or any other input value.

Why: to see which kinds of PDFs fail so they can be fixed, and which documents people bring so the tool is built for them.

Opt out: uncheck "Share anonymous run statistics" (shareAnonymousStats: false). Nothing is sent then.

Links are shared only if you ask. If you check "Share links of PDFs that failed" (shareFailingLinks: true), the links of the PDFs where no table was found, or that failed, are added to the record, at most 20 per run, with everything after ? or # removed (also when written as %3F or %23), any login in the link removed, and any path parameters removed too (for example the ;jsessionid=... some servers add, also when the semicolon is written %3B). Links that were refused are never sent.

Kept for 180 days, then deleted.

Apify also has an account setting, "Share run data with developers". It is separate from this record and is yours to switch on or off: when it is on, the developer can see your runs' inputs and outputs. The developer's own test runs send nothing. To ask about any of this, open the Issues tab of this Actor.

Limits

• Born-digital PDFs only. Scanned PDFs (images of pages) have no text to read, so they return no-tables. There is no OCR.
• The table needs some rules: a grid, or lines above and below the header. A table with no lines at all is not found. That is deliberate, so paragraphs of prose never come back as fake tables.
• Only public links. Links to private or internal network addresses are refused, at every redirect, and the download connects to the address that was checked. Plain http links work, because many public PDFs are still served that way.
• A download that takes more than two minutes is given up (a failed record, not charged), however slowly the server sends its response headers or the file, and so is a PDF over 50 MB. A server that sends the file compressed (a Content-Encoding) is refused.
• If Apify restarts or moves a run, the PDFs it had already finished are skipped, not read or charged again. The one PDF that was in progress is read again, and if the restart came while its rows were being saved or just before its charge, its rows can appear twice in the dataset; it is charged once, when it finishes. A charge can repeat only if the restart falls in the instant between a PDF's charge and the saving of its progress record, or if that record could not be saved (the run log says so).
• Very large PDFs are limited by maxPages rather than by a time limit.
• Column names come from the table's heading. A heading that spans several columns is split across them, and a heading printed outside the table's lines is not read: the columns are then named column_1, column_2, ... Footnote markers stay on labels (Apparel1). A table whose first data rows hold only codes other than the common ones ((D), (NA), --, n/a), such as * or S, can have those rows read into its heading.
• Every table carries a confidence score. For a table drawn as a grid it is how well two independent readings agree (the cell borders, and the whitespace between words). For a table with lines only above and below its heading there is one reading, and the score is the share of words that fall cleanly inside one column. Below 0.9, check the table by eye.
• A page that cannot be read is skipped, never silently: the PDF's other tables still come back, and a partial record names the skipped pages.

Accuracy

Real PDFs. 20 public releases from BLS, the Census Bureau, the Federal Reserve, the IRS and the EIA (October 2026), first 50 pages each: 261 tables found, 28 of them (11%) scored below 0.9, most of those IRS worksheet forms (numbered instruction lines with amount boxes) and a few tables whose cells run to paragraphs. The EIA weekly petroleum report yields none, because its tables have no lines. Each PDF took under 10 seconds on a desktop PC.

Ten tables picked at random, their first rows compared cell by cell with the page: all ten were real tables, and every value matched. Four earlier samples found faults that were fixed, among them row labels cut off where a table has no outer rule, merged columns, numbers split by a stray space, and prose or instructions returned as tables. Two independent code reviews then built about 35 more shapes that the fixes had broken or never handled (headings of years or column numbers, suppressed values such as (D), grouped rows with a blank first cell, fee schedules with long labels, sparse grids, white headings on rounded or spot-colour bands, tables drawn as white cells on a dark backing, two same-width tables on one page with prose, a caption or only space between them, a contents page). Each one is now a test case, and real pages are pinned where the shape came from one.

Generated PDFs. 89 PDFs with exact answer keys, over 40 table shapes and 7 pages that hold no table (prose, a contents page, a cover box, a worksheet): cell accuracy 100% on every shape, every heading kept as the heading, and no table found where there is none.