PDF to Text and OCR: Extract Text from PDFs and Scans

One row per page: text, method used, word count and, for scanned pages, OCR confidence. Pages of digital PDFs are read from their text layer, which is exact and fast. Scanned pages and images go through Tesseract OCR, in English, French, German, Spanish, Italian, Portuguese or Dutch. You pay per page, never per document, and failed pages are free.

What it does

• Digital PDFs: the text layer is extracted page by page with poppler (pdftotext), line breaks kept.
• Scanned PDFs and images (PNG, JPEG, TIFF, BMP, GIF, WebP): each page is rendered at the resolution of its scan (150 to 300 dpi) and read by Tesseract 5. With ocrMode set to auto, OCR runs only on pages without a text layer, so a PDF that mixes typed and scanned pages costs OCR only for the scanned ones. A multipage TIFF gives one row per page, like a PDF.
• Files from your Apify account: pick a key-value store in keyValueStoreId and every file in it is read. The Actor gets read access to that store only.
• The text is returned as it is. Nothing is extracted from it: no names, no amounts, no IBAN, no passport or ID card zone. Use the text in your own pipeline (search index, translation, LLM, archive).

Use cases

• Make a set of scanned reports, books or archives searchable.
• Feed the text of PDF documents to an LLM or a RAG index, page by page with the page number kept.
• Check which PDFs of a collection are scans without a text layer (method = ocr).
• Extract the text of public documents (laws, tenders, annual reports, tax guides) in bulk.

Not for identity documents or health records

Do not send ID cards, passports, driving licences, payslips, bank statements or medical documents. The Actor does not recognize or extract any field, and documents you send may contain names of people: you are responsible for having the right to process them. Downloaded files are deleted as soon as their pages are read, and nothing is kept outside the dataset of your run.

Input

Field	Description	Default
sources	Public URLs of PDF files or images, one per line. The type is read from the file content
keyValueStoreId	Optional key-value store of your account holding the files (every record except INPUT)
ocrMode	auto: OCR only for pages without a text layer. always: OCR every page. never: text layer only	auto
ocrLanguages	Languages of the scanned text, up to 3: eng, fra, deu, spa, ita, por, nld	["eng"]
maxPages	Page rows written in the run, over all files, failed and blank pages included	1000 (up to 100,000)
maxPagesPerFile	Only the first pages of longer files are read	500 (up to 5000)

Up to 1,000 files per run, 100 MB per file and 50 megapixels per image page (about 7,000 x 7,000 pixels); larger images are refused with a free error row. A file whose download takes more than 5 minutes is reported as too slow. Only public addresses are fetched: URLs that point to a private or local network are refused.

Example:

{
  "sources": [{ "url": "https://www.irs.gov/pub/irs-pdf/p15.pdf" }],
  "ocrLanguages": ["eng"],
  "maxPagesPerFile": 20
}

Output

One row per page. Real rows from a run on IRS Publication 15 and on a scanned page of Jules Verne's De la Terre à la Lune (1865 edition, public domain) on 28 October 2026, text shortened here:

[
  {
    "sourceUrl": "https://www.irs.gov/pub/irs-pdf/p15.pdf",
    "fileName": "p15.pdf",
    "fileType": "pdf",
    "page": 2,
    "pageCount": 59,
    "method": "text-layer",
    "ocrLanguages": null,
    "ocrConfidence": null,
    "wordCount": 973,
    "charCount": 6331,
    "text": "the employer and employee. The social security wage\nbase limit is $184,500.\nThe Medicare tax rate is 1.45% each for the employee\nand employer, unchanged from 20...",
    "error": null
  },
  {
    "sourceUrl": "https://upload.wikimedia.org/wikipedia/commons/thumb/a/ae/Verne_-_De_la_Terre_%C3%A0_la_lune.djvu/page19-1920px-Verne_-_De_la_Terre_%C3%A0_la_lune.djvu.jpg",
    "fileName": "page19-1920px-Verne_-_De_la_Terre_à_la_lune.djvu.jpg",
    "fileType": "image",
    "page": 1,
    "pageCount": 1,
    "method": "ocr",
    "ocrLanguages": "eng+fra",
    "ocrConfidence": 92.4,
    "wordCount": 343,
    "charCount": 2222,
    "text": "EFFET DE LA COMMUNICATION BARBICANEF. 15\n\n—————\n\nhaletantes; puis il se fit un moment de silence, semblable à ce calme\nprofond qui précède les coups de tonnerre. E...",
    "error": null
  }
]

A file that cannot be read gives one row with page = null and the reason in error (HTTP 404, not a PDF or an image, password-protected PDF, larger than 100 MB). A page that fails gives a row with its page and error. These rows are free.

Accuracy

Measured on 28 October 2026 on 20 pages, with the share of the words of a reference text found in the OCR output:

Pages	Words right
10 pages of digital PDFs (IRS Publication 15 with 2 table pages, French Constitution) rendered as 200 dpi scans	97.7 %
5 book scans from Wikisource (English and French), compared with the proofread transcription	97.8 %
The same kind of scans degraded to about 150 dpi and tilted by 2.5 to 3 degrees	97.4 %

Over the 10 real scans (5 clean, 5 degraded) the score is 97.6 %. Expect less on photos taken with a phone, handwriting (not supported), very small print and complex tables, whose cells come out as lines of text without the grid.

Pricing

• $0.0005 per page read from the text layer ($0.50 per 1,000 pages).
• $0.004 per page read by OCR ($4 per 1,000 pages).

Pages that fail and blank pages are free. After 10 pages in a row where OCR finds no text, the Actor stops reading that file and says so in an error row, so a blank or unreadable file does not use up run time. There is no fee per file or per run start, and volume discounts apply on higher Apify plans. Set a maximum cost per run in the Apify Console: the Actor checks it before each OCR page and stops cleanly, without charging pages it did not read.

Run time

The default memory of 2 GB reads about 4 to 6 OCR pages per minute; text-layer pages take a fraction of a second. OCR reads one page at a time on one CPU core, so more memory does not make it faster: 2 GB is enough, and the maximum is 4 GB. The run stops one minute before its timeout (2 hours by default) and ends normally with the pages read so far. If the platform moves the run to another server, it resumes there without charging the same pages twice.

Integrations and recipes

• Apify MCP server: let an AI assistant call this Actor to read a PDF or a scan and get the text page by page.
• Make, Zapier, n8n: start a run from a new file in Google Drive or an e-mail attachment URL, then read the dataset items to index, translate or summarize the text.

FAQ

Do I need an API key? No. OCR runs inside the Actor with Tesseract; nothing is sent to a third-party service.

Which languages? English, French, German, Spanish, Italian, Portuguese and Dutch, up to 3 per run.

Does it keep my files? No. Downloaded files are deleted as soon as their pages are read.

Does it read handwriting, invoices or ID cards? No. Handwriting is not supported, and no field is extracted (no amounts, no names). Do not send identity documents, payslips, bank statements or health records.