Extract Text From PDF: text per page for LLMs and RAG

Extract text from PDF files by URL, up to 200 files per run: one row per page (or one per file) with the text, page count, title, author, dates, producer and file size, plus whether the page has a text layer. Optional: tables as cells and as Markdown. Pay per page delivered.

Why this scraper

• Only delivered results are charged. Inputs that fail come back as clear error records at no cost.
• Pure Python parsing, no browser and no OCR engine: a 15-page research paper comes back page by page in a few seconds. Measured on the platform: two files, six pages, for a hundredth of a cent.
• One row per page (or one per file), with the character count and a hasTextLayer flag per page, so scanned pages show up as exactly that instead of silently empty. The document's own metadata rides along: title, author, subject, creation and modification dates as ISO timestamps, producer, file size, encryption flag. Every URL is checked against its host's robots.txt before the file is fetched, and files above 50 MB are refused for free.
• Tables as Markdown, at the same price. Switch on extractTables and every page row also carries its ruled tables, as clean rows of cells and as a Markdown table you can drop straight into an LLM prompt or a RAG chunk. Merged cells and headers spread over several lines are folded into one header row and real columns, instead of the dozens of empty columns a raw table parser returns.

Who this is for

Paste direct PDF links in urls (up to 200 per run). Leave mergePages off for one row per page, or switch it on for one row per file with all pages joined (that row is charged as the number of pages it holds, the same as per page). maxPagesPerFile caps the cost per file. Every row carries the URL, file name, page number, page count, text, character count, text-layer flag, title, author, subject, dates, producer, file size and encryption flag.

Who this is not for

This reads the text layer; a scanned PDF (images of pages) comes back with empty text and hasTextLayer false on every page, because there is nothing to read without OCR. Title, author and subject are whatever the producing software wrote into the file, and most files carry no title or subject at all (a research paper from LaTeX has a producer but no author field). Reading order follows the PDF's internal order, which can differ from the visual layout in multi-column documents. Password-protected files and pages that pypdf cannot parse are reported, not charged. Table extraction finds tables drawn with lines; a table laid out with spaces alone, without ruling, stays in the page text and does not appear in tables. That includes most tables in typeset research papers, which use only a few horizontal rules: their content is in text, not in tables.

Input fields

Field	Type	Required or default	What it does
urls	list of text	required	One URL per row, up to 200. Direct links to PDF files.
mergePages	true/false	false	Off: one row per page. On: one row per file with all pages joined; that row is charged as the number of pages it holds.
maxPagesPerFile	number	200	Cost ceiling per file, from the first page.
extractTables	true/false	false	On: every page row also carries the ruled tables on that page, as rows of cells and as a Markdown table. Slower on long files; the price stays per page.

Input example

{
    "urls": [
        "https://www.w3.org/WAI/ER/tests/xhtml/testfiles/resources/pdf/dummy.pdf"
    ],
    "mergePages": false,
    "maxPagesPerFile": 200
}

Output example

Field	Type	What it holds	Example
url	text	The page or file this row is about.	https://arxiv.org/pdf/1706.03762
fileName	text	The file name from the link or from the response header.	1706.03762
page	number	The page number this row comes from; one row per page.	1
pageCount	number	How many pages the PDF holds in total.	15
text	text	The plain text, without markup, ready to search or feed to a model.	1 Introduction Recurrent neural networks, long short-term...
characters	number	How many characters of text this page yielded.	2855
hasTextLayer	true/false	True when the page carries real text; false means a scan, where you need OCR instead.	True
title	text	The title in the PDF's own properties; often empty or the name of the program that made it.
author	text	The author the source names; empty when it names none.
subject	text	The subject in the PDF's properties.
createdAt	text	When the PDF was created, according to its own properties.	2024-04-10T21:11:43Z
modifiedAt	text	When it was last changed, according to its own properties.	2024-04-10T21:11:43Z
producer	text	The program that produced the PDF, which often says how it was made.	pdfTeX-1.40.25
fileBytes	number	The size of the file in bytes.	2215244
encrypted	true/false	True when the PDF is protected; a protected file yields no text.	False
tables	list
tablesMarkdown	text

Error codes: INVALID_URL, ROBOTS_DISALLOWED, NOT_A_PDF, FETCH_FAILED, TOO_LARGE, BLOCKED.

One delivered row looks like this:

{
  "url": "https://www.w3.org/WAI/ER/tests/xhtml/testfiles/resources/pdf/dummy.pdf",
  "fileName": "dummy.pdf",
  "pageCount": 1,
  "title": null,
  "author": "Evangelos Vlachogiannis",
  "subject": null,
  "createdAt": "2007-02-23T17:56:37",
  "modifiedAt": null,
  "producer": "OpenOffice.org 2.1",
  "fileBytes": 13264,
  "encrypted": false,
  "page": 1,
  "text": "Dummy PDF file",
  "characters": 14,
  "hasTextLayer": true,
  "status": "ok"
}

Related actors from steadydata

• sitemap-urls: find every PDF a site lists in its sitemap
• url-status-redirect-checker: check which PDF links still answer
• google-patent-details: patent texts without the PDF
• Used together with sitemap-urls and webpage-to-markdown in a recipe: Feed a RAG pipeline with clean page text.

Pricing

Pay per event: one page-extracted event per delivered result. No charge for inputs that fail, no separate platform-usage surcharge.

Free Apify plan: this actor delivers up to 25 rows per run for accounts on the Apify free plan, and then stops with a message. That limit is set by us, not by Apify. It exists so the actor keeps paying for itself for the people who do pay. Any paid Apify plan runs it at full size, billed per delivered row, with failed rows never charged.

Reviews: if this actor saves you time, a short review on this page is the one thing that helps most. Ratings are what other buyers look at first, and we have no other way to ask.

FAQ

Does it do OCR? No. It extracts the existing text layer; the hasTextLayer flag tells you which pages would need OCR.

Can I get the whole document as one text? Yes, switch on mergePages; the joined row is charged per page it holds, so the price is the same as one row per page.

What about very large files? Files above 50 MB are refused as a free error row; use maxPagesPerFile to cap long documents.

Where do the URLs come from? Any direct PDF link: your own site, the Sitemap URLs actor below, a search result. A page that returns HTML instead of a PDF is a free NOT_A_PDF error row.

Do I pay for a file that fails? No. Only delivered pages (or merged files) are charged.

Is personal data collected? page is 1-based, or 0 for a merged file; characters counts the cleaned text; createdAt and modifiedAt are the PDF info dates as ISO timestamps without a time zone.

What happens when the source changes? Sources change from time to time; that is the nature of this work. The actor is monitored daily and fixed fast, and while it is broken you are not charged, because only delivered results cost anything.

Can I call it from my own code, like an API? Yes. Every run is available through the Apify API: start it with one HTTP request, then read the results as JSON, CSV or Excel from the dataset endpoint, or use the official Python and JavaScript clients. The input is the same JSON as in the input example above, so a scheduled script, a Make or Zapier step or an AI agent can call it without any changes.