Hebrew PDF & OCR Text Extractor (RTL-correct)

Turn Hebrew PDFs, scanned documents, Word files, images and web pages into clean, correctly ordered text, ready for search, spreadsheets, translation or an LLM.

Most PDF and OCR tools were built for left-to-right languages. On Hebrew they quietly break things. Punctuation jumps to the wrong end of the line, brackets come out backwards, numbers and English words land in the wrong place, and sometimes the whole line is reversed. This Actor was built and tested on real Israeli documents (Bank of Israel reports and Tel Aviv University and Hebrew University regulations) to fix exactly those problems.

	Typical extractor output	This Actor
Section number	.2.3 תואר "דוקטור לפילוסופיה" ( ).Ph.D	2.3. תואר "דוקטור לפילוסופיה" (Ph.D.)
English in brackets	קידום תשלומים מיידיים (,)Faster Payments	קידום תשלומים מיידיים (Faster Payments),
Mirrored brackets (Acrobat PDFMaker)	חילופי סטודנטים )פטורים(	חילופי סטודנטים (פטורים)
Percentages	מע"מ (%17)	מע"מ (17%)
Scanned line read left-to-right by OCR	.2.0 בגרסה PDF-ב Microsoft Word המערכת	המערכת Microsoft Word PDF-ב בגרסה 2.0.
Legacy Windows-1255 text	äñëí ùéøåú	הסכם שירות
Hidden bidi characters	invisible U+202B…U+202C everywhere	removed

What it does

• 📄 PDF with a text layer is read directly. It's fast, exact and costs the least.
• 🖨️ Scanned PDFs and photos are read with Tesseract 5 OCR using the official Hebrew and English models. It decides page by page, so a PDF that mixes typed and scanned pages just works.
• 📝 DOCX (Word) files keep their headings, numbered and bulleted lists, and tables.
• 🌐 Web pages: the main content only, without menus, footers or scripts. Old ISO-8859-8 "visual Hebrew" pages are reversed into logical order.
• 🔤 Text files in UTF-8, UTF-16 or Windows-1255, with automatic encoding repair.
• ↔️ RTL repair: bidi control characters are removed, punctuation and bracket order is repaired, mirrored brackets are detected per document, and reversed visual text is detected and flipped.
• 🧾 Output is plain text or Markdown (headings, lists, tables, re-joined paragraphs), with one row per document or one row per page, plus optional .txt / .md files.
• 🏷️ Each row also has the language (he, en, he+en, ar, ru …), the text direction, the share of each script, OCR confidence and a list of the fixes applied.
• 💸 Safe by design: a dry run shows the cost first, there is a max-pages cap, the max run cost is respected, and failed documents are never charged.

Other OCR languages are available too: Arabic, Russian, Yiddish, French, German, Spanish, Italian and Portuguese, and combinations such as heb+eng+ara or heb+eng+rus.

How to use it

1. Paste one or more document links into Document URLs, upload files from your computer, or pick an Apify key-value store whose records should be read (for example, files saved by another Actor). Links from Google Drive, Google Docs, Dropbox and GitHub are converted to direct downloads automatically.
2. Keep OCR languages on Hebrew + English unless your documents need more.
3. Click Start. Each document becomes one dataset row (or one row per page), which you can export as JSON, CSV, Excel, or read through the API.

💡 Not sure how much a big batch will cost? Turn on Dry run first. It counts text pages and scanned pages for each file and gives an estimated price, and nothing is charged.

Input example

{
  "documents": [
    { "url": "https://www.example.co.il/reports/annual-report-2026.pdf" },
    { "url": "https://drive.google.com/file/d/FILE_ID/view" }
  ],
  "ocrLanguages": "heb+eng",
  "ocrMode": "auto",
  "outputFormat": "markdown",
  "maxPagesPerDocument": 50,
  "oneItemPerPage": false
}

Output example

This is a real result (shortened) for a Bank of Israel report:

{
  "url": "https://boi.org.il/media/flpbnbte/…במזומן.pdf",
  "status": "ok",
  "detectedType": "pdf",
  "method": "text-layer",
  "pageCount": 16,
  "pagesProcessed": 3,
  "textPages": 3,
  "ocrPages": 0,
  "chars": 4745,
  "words": 777,
  "language": "he",
  "direction": "rtl",
  "scripts": { "hebrew": 0.997, "latin": 0.003 },
  "fixesApplied": ["bidi-controls-removed", "rtl-punctuation-reordered"],
  "text": "<!-- page 1 -->\n\nסקירת ההתפתחויות\nבשימוש במזומן\n\nבנק ישראל\n…\n\n<!-- page 2 -->\n…\nפתח דבר\nמאת מנהלת מחלקת המטבע\nהתמורות הטכנולוגיות הרבות שנכנסו לחיינו בשנים האחרונות הרחיבו את מגוון האפשרויות העומדות בפני הציבור בכל הקשור להתנהלותו בעולמות התשלומים. …",
  "billedTextUnits": 3,
  "billedOcrPages": 0,
  "warnings": ["Only the first 3 of 16 pages were processed (maxPagesPerDocument)."],
  "error": null
}

A failed document still gets a row that explains what went wrong, and it costs nothing:

{ "url": "https://example.com/missing.pdf", "status": "failed", "chars": 0,
  "error": { "code": "NOT_FOUND", "message": "The server returned 404 Not Found." } }

Error codes: NOT_FOUND, ACCESS_DENIED, BLOCKED_OR_NOT_A_DOCUMENT, TIMEOUT, NETWORK_ERROR, HTTP_ERROR, FILE_TOO_LARGE, UNSUPPORTED_TYPE, PDF_PASSWORD_PROTECTED, PDF_DAMAGED, IMAGE_UNREADABLE, NO_TEXT, OCR_FAILED, BUDGET_EXHAUSTED.

Pricing

You pay only for pages that were actually read:

Event	What counts as one
Text page	A PDF page read from its text layer, or each started 3,000 characters of a DOCX, web page or text file
OCR page	A scanned PDF page or one image (or one frame of a multi-page TIFF) read with OCR
Proxy megabyte	Only if you turn on a proxy: each MB downloaded through it

The current prices are shown on the Pricing tab. There is no start fee. Failed documents, empty pages and dry runs are free. Max pages per document and the run's maximum cost limit are always respected. When the limit is reached, the Actor stops cleanly and tells you so.

Israeli government websites (gov.il, knesset.gov.il)

Many Israeli government sites only serve visitors from Israel. Everyone else gets an "Unavailable" web page instead of the PDF. The Actor detects this and returns BLOCKED_OR_NOT_A_DOCUMENT with a hint instead of extracting rubbish. To read these files, open Proxy configuration, choose Residential proxies and set the country to Israel (IL).

Use it from code

curl -X POST "https://api.apify.com/v2/acts/vantum~hebrew-document-reader/run-sync-get-dataset-items?token=YOUR_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"documents":[{"url":"https://example.co.il/file.pdf"}],"outputFormat":"markdown"}'

from apify_client import ApifyClient

client = ApifyClient("YOUR_TOKEN")
run = client.actor("vantum/hebrew-document-reader").call(run_input={
    "documents": [{"url": "https://example.co.il/file.pdf"}],
    "oneItemPerPage": True,  # one row per page, handy for RAG chunking
})
for page in client.dataset(run["defaultDatasetId"]).iterate_items():
    print(page["page"], page["text"][:80])

It works with Make, Zapier, n8n, LangChain, LlamaIndex and the Apify MCP server like any other Apify Actor.

Tips

• Use Markdown output for LLMs and RAG. Headings, lists and tables are kept, and hard-wrapped PDF lines are joined back into paragraphs.
• Use One dataset row per page for search indexes or page citations.
• If a PDF's text layer is broken (wrong letters or gibberish), set When to use OCR to Always.
• Use Never OCR when you only want the cheap text layer and don't care about scanned pages.
• Remove niqqud makes vocalised texts (poetry, prayer books, children's books) easier to search.
• Memory: OCR uses one CPU core per 4 GB. 2 GB is fine for most jobs. Give it 8–16 GB for big scanned batches so pages are OCRed in parallel.

Limitations

• Handwriting is not supported. Tesseract reads printed text.
• On scanned pages, a line that mixes Hebrew and English words is the hardest case for any OCR engine. Word order is repaired, but individual mixed-script words can be misread more often than pure Hebrew or pure English. Text-layer PDFs, DOCX and web pages are not affected.
• Tables in text-layer PDFs come out one cell per line. Turn on Preserve PDF layout to keep the columns.
• These formats are reported as UNSUPPORTED_TYPE with a hint on how to convert them: legacy .doc, .xlsx, .pptx, .odt, ZIP archives and HEIC photos.
• Password-protected PDFs are reported as PDF_PASSWORD_PROTECTED.
• Pages that need JavaScript to render (single-page apps) may have no text in their HTML. Link to the document file itself instead.

FAQ

Is my data stored? The files are processed inside your own Apify run, and the text is saved only to your run's dataset and key-value store. Nothing is sent to third-party AI or OCR APIs. Everything runs on open-source software inside the container.

How accurate is the OCR? On 300-dpi renders of real Hebrew reports and regulations, OCR output matched the original text layer at 95–99% word-sequence similarity in our tests (pages full of dot leaders, like tables of contents, score lower). Each page reports Tesseract's confidence in ocrConfidence, so you can flag weak scans.

Does it handle Arabic too? Yes. The RTL repair works for Arabic, and ara / heb+ara OCR models are included.

---

בעברית

חילוץ טקסט נקי ובסדר הנכון מקבצי PDF, סריקות, Word, תמונות ודפי אינטרנט בעברית.

רוב כלי ה-PDF וה-OCR נבנו לשפות שנכתבות משמאל לימין, ובעברית הם משבשים את הטקסט. סימני פיסוק קופצים לצד הלא נכון, סוגריים מתהפכים, ומספרים ומילים באנגלית זזים ממקומם. הכלי הזה נבנה ונבדק על מסמכים ישראליים אמיתיים, והוא מתקן בדיוק את הבעיות האלה.

• קבצי PDF עם שכבת טקסט נקראים ישירות, ומסמכים סרוקים ותמונות עוברים זיהוי תווים (OCR) בעברית ובאנגלית. ההחלטה מתקבלת לכל עמוד בנפרד.
• קבצי Word ‏(DOCX) נקראים כולל כותרות, רשימות וטבלאות. מדפי אינטרנט נשלף רק התוכן עצמו, ודפי "עברית ויזואלית" ישנים מתוקנים.
• קידוד ישן (Windows-1255) מתוקן אוטומטית, ואפשר גם להסיר ניקוד וטעמים.
• הפלט הוא טקסט רגיל או Markdown, בשורה אחת לכל מסמך או לכל עמוד, ועם זיהוי שפה וכיוון כתיבה.
• משלמים רק על עמודים שנקראו בפועל. מסמכים שנכשלו והרצת ניסיון (Dry run) לא עולים כלום.
• אתרי ממשלה (gov.il, knesset.gov.il) חוסמים גלישה מחוץ לישראל. כדי לקרוא מהם, הפעילו פרוקסי Residential עם מדינה ישראל (IL).

Open-source components

The Actor uses Tesseract OCR and the official tessdata models (Apache-2.0), Poppler utilities (GPL, called as separate programs), python-docx (MIT), Beautiful Soup (MIT), HTTPX (BSD), Pillow (MIT-CMU) and the Apify SDK (Apache-2.0). See THIRD_PARTY_LICENSES.md.