Extract text from PDFs in bulk, with OCR for scanned pages

Do it yourself: pdftotext and Tesseract

Poppler's pdftotext reads the text layer. Tested on 2026-10-12 with the Bitcoin white paper:

curl -sL -o bitcoin.pdf https://bitcoin.org/bitcoin.pdf
pdfinfo bitcoin.pdf | grep -E "Pages|Producer"     # Pages: 9, Producer: OpenOffice.org 2.4
pdftotext -layout bitcoin.pdf - | head -3          # "Bitcoin: A Peer-to-Peer Electronic Cash System"
pdftotext bitcoin.pdf - | wc -w                    # 3568 words

A scanned page has no text layer, so pdftotext returns nothing. Render it and run Tesseract. OCRmyPDF's skew.pdf test file:

pdftotext skew.pdf - | wc -w                      # 0 words: no text layer
pdftoppm -r 300 -gray -png skew.pdf skew           # render at 300 dpi
tesseract skew-1.png - | head -2                   # "The LinnSequencer" / "32 Track MIDI Sequence Recorder"
tesseract skew-1.png - | wc -w                     # 729 words

In Python, pypdf or pypdfium2 do the same. What a script has to add for a real batch:

• Deciding per page whether OCR is needed (OCR-ing everything is slow and makes text worse on clean pages).
• Telling apart broken downloads, HTML error pages served as .pdf, encrypted files and empty scans.
• Cutting text into chunks at paragraph or sentence boundaries with overlap for embeddings.
• Reading PDF metadata dates such as D:20240306081813-05'00' into ISO dates.
• Installing Tesseract and its language packs wherever the job runs.

Extract text from many PDFs without code

1. Open PDF Text Extractor: PDF to Text, Pages & OCR on Apify.
2. Paste direct PDF links, up to 10,000.
3. Choose One row per PDF or One row per page, and optionally a Chunk size (for example 1,500 characters with 200 overlap) for RAG.
4. Leave OCR scanned pages on and pick the languages (English, German, French, Spanish, Italian, Portuguese, Dutch).
5. Click Start and export JSON, CSV or Excel, or call it from your pipeline through the API.

Text comes from PDFium, the PDF engine in Chrome. Only pages with fewer than 20 non-space characters are rendered at 300 dpi and read with Tesseract 5. A row from a real run on 2026-10-12, text shortened:

{
  "url": "https://www.irs.gov/pub/irs-pdf/fw9.pdf",
  "status": "ok",
  "pageCount": 6,
  "pagesExtracted": 6,
  "pagesOcr": 0,
  "wordCount": 6516,
  "title": "Form W-9 (Rev. March 2024)",
  "subject": "Request for Taxpayer Identification Number and Certification",
  "producer": "Designer 6.5",
  "creationDate": "2024-03-06T08:18:13-05:00",
  "pdfVersion": "1.7",
  "text": "Form W-9\n(Rev. March 2024)\nRequest for Taxpayer\nIdentification Number and Certification …"
}

Worked examples from our test runs

The local Tesseract run above found 729 words on the same scan, so the actor's OCR matches a plain Tesseract call; the gain is doing it only where needed and across thousands of files.

What each row contains

Honest limits

• Tables become text lines, not cells, and no Markdown headings are generated. For table extraction, use a layout-aware tool such as Camelot.
• Multi-column layouts usually come out column by column, following the PDF's text objects.
• Encrypted PDFs are reported as encrypted and skipped; handwriting is not supported; OCR quality depends on the scan.
• Files must be at a public URL (no uploads, no logins); 50 MB per file by default, 200 MB at most.
• Downloads follow each site's robots.txt and do not get around bot protection.

All input fields

Generated from the actor's live input schema (build 1.0.4).

Call it from code

Start a run and get the results in one HTTP request with the Apify API (your token is under Console, Settings, API & Integrations):

curl -X POST "https://api.apify.com/v2/acts/adamhaleworks~pdf-text-extractor-ocr/run-sync-get-dataset-items?token=$APIFY_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"urls":["https://bitcoin.org/bitcoin.pdf"],"chunkSize":1000}'

The run's Output tab offers 2 ready-made views; every view exports to CSV, Excel, JSON or XML:

• Overview: url, status, pageCount, pagesOcr, wordCount, title, author, creationDate, text, warning, error
• Pages (one row per page mode): url, page, pageCount, pageOcr, pageTextLength, pageText

What it costs

At the planned prices, the 3-PDF default run would cost $0.006 and the scanned page another $0.003. Failed downloads, non-PDFs, encrypted or broken files, PDFs with no text and OCR that finds nothing would not be charged.

FAQ