PDF OCR: Scanned PDF & Image to Text, Searchable PDF

PDF OCR and image OCR in one Actor, for developers building RAG and AI-agent pipelines, and for anyone who has to dig text out of scans: give this Actor links to images or scanned PDFs (PNG, JPEG, TIFF incl. multi-page, WebP, GIF, BMP, PDF) and it returns the text of every page, RAG-ready chunks with page numbers, optional word boxes with confidence, and an optional searchable PDF (grayscale page images with an invisible text layer).

From $1.20 per 1,000 OCR pages ($2 on the Free plan, less on paid plans). PDF pages that already have text are read without OCR for $0.30 per 1,000, and blank pages and failures are free. OCR runs with Tesseract inside the Actor: your files are not sent to a third-party OCR API.

What you can use it for

• RAG and AI agents: scanned contracts, letters, receipts and old reports become text with page numbers, ready for a vector database.
• Searchable archives: turn a folder of scans into searchable PDFs that open in any PDF viewer.
• Mixed PDFs: a PDF where only some pages are scans is handled in one run; text pages are copied exactly and charged at the lower price.
• Data entry: word boxes (position and confidence) let you pick out fields such as totals and dates.

Output example

The prefilled example (phototest.tif from the Tesseract test set), Output format = Text (one row per file):

{
  "url": "https://raw.githubusercontent.com/tesseract-ocr/test/main/testing/phototest.tif",
  "fileName": "phototest.tif",
  "fileType": "tiff",
  "pageCount": 1,
  "pagesProcessed": 1,
  "pagesOcr": 1,
  "pagesTextLayer": 0,
  "pagesBlank": [],
  "pagesFailed": [],
  "languages": "eng",
  "meanConfidence": 96.4,
  "text": "This is a lot of 12 point text to test the\nocr code and see if it works on all types\nof file format.\n\nThe quick brown dog jumped over the\nlazy fox. ...",
  "charCount": 285,
  "pages": [
    { "page": 1, "method": "ocr", "confidence": 96.4, "rotation": 0, "text": "This is a lot of 12 point text to test the\n..." }
  ],
  "searchablePdfUrl": null,
  "error": null,
  "processedAt": "2026-10-21T03:49:28.376303+00:00"
}

The same page turned sideways (phototest-rotated-L.png) comes back with the same text and "rotation": 90. A two-page PDF that already has text comes back with "method": "text-layer" on both pages and is charged as two text-page events.

With Include word boxes, each page also has words, e.g. { "text": "quick", "confidence": 95.9, "left": 102, "top": 194, "width": 75, "height": 30, "block": 1, "line": 1 }. Positions are in pixels of the original image (for PDF pages: pixels at 300 DPI), measured on the page after its orientation was fixed.

How to use

1. Put the files somewhere the Actor can download them (a public link, a pre-signed S3 or Cloud Storage URL, a direct Dropbox link, an Apify key-value store record) and list the links under Image or PDF URLs.
2. Pick the Languages of the text (English by default, up to 3 at once).
3. Optionally turn on Create a searchable PDF or Include word boxes, or choose RAG chunks as the output format.
4. Press Start and wait for the run to finish (seconds for a few images). Each file is one row in the run's dataset, which you can export in several formats or fetch over the API.

The prefilled example reads phototest.tif, a test image from the Tesseract project, and finishes in under a minute.

Input

Field	What it does
Image or PDF URLs	Links to images (PNG, JPEG, TIFF, WebP, GIF, BMP) or PDFs. The type is detected from the file content. Up to 50 MB per file. You can also link a text file that lists URLs (up to 1,000).
Languages	eng (default), deu, fra, spa, por, ita, nld, pol, tur, rus, ukr, jpn, chi_sim, kor. Up to 3 at once.
Output format	text (one row per file with a pages list) or chunks (one row per RAG chunk with pageStart / pageEnd).
Create a searchable PDF	Saves one PDF per file in the run's key-value store; the link is in searchablePdfUrl. No extra charge.
Include word boxes	Adds every word with its confidence (0-100) and pixel position to each page.
Fix page orientation	On by default: pages that are sideways or upside down are turned before OCR.
OCR pages that already have text	Off by default. Turn on if a PDF's existing text layer is wrong.
Max pages per file	0 = all, up to 500.
Chunk size / Chunk overlap	For RAG chunks: maximum characters per chunk (default 2,000) and characters repeated from the previous chunk (default 200).

How it works

1. Each file is downloaded (at most one request per second to the same host; busy servers are retried twice).
2. PDF pages with a usable text layer are read directly. Other pages are rendered at 300 DPI; images are converted to grayscale, and small images (screenshots, thumbnails) are enlarged so small letters can be read.
3. Tesseract detects the orientation, then recognizes the text (LSTM engine). Each page gets a method: ocr, text-layer, blank or failed.
4. With a searchable PDF, OCR'd pages are rebuilt as image + invisible text and other pages are copied as they are, in the original order.

Pricing

Pay per event:

Event	Free plan	Bronze	Silver	Gold and above
ocr-page: a page read by OCR that has text	$0.002	$0.0018	$0.0015	$0.0012
text-page: a PDF page read from its own text layer	$0.0003	$0.0003	$0.0003	$0.0003

Plus Apify's tiny start fee of $0.00005 per GB of memory per run (the default 2 GB counts as two). Platform usage is included.

Examples on the Free plan: a 10-page scanned PDF costs $0.02. A 30-page PDF where 3 pages are scans costs 3 × $0.002 + 27 × $0.0003 = $0.0141. 1,000 receipts photographed with a phone cost $2.

Blank pages, pages OCR cannot process and files that cannot be downloaded are reported and not charged. If you give the run a spending limit, the Actor reads only as many pages as the limit covers and skips the remaining files.

Limits

• OCR quality depends on the scan: clear printed text at 200-300 DPI works best. Handwriting is not supported. Low-contrast photos, decorative fonts and very small text give lower confidence (see meanConfidence and the per-word confidence).
• Layout is returned as text lines and paragraphs; tables come out as text lines, not Markdown tables.
• Up to 50 MB per file, 500 pages per file and about 100 megapixels per image (larger images are reported as an error). For very long documents the row keeps the full text and the per-page methods; the per-page texts are dropped (pageTextsOmitted) so the row fits the 9 MB dataset limit. OCR takes roughly a few seconds per page with the default 2 GB of memory; give the run more memory (4 GB = one full CPU core, 8 GB = two) to process pages in parallel.
• Password-protected PDFs are not supported. Vertical Japanese and Chinese text is not supported (horizontal only).
• The searchable PDF is built from grayscale page images, so colors are not kept. Keep the original if you need it.
• A PDF page that has only a short text layer (under 200 characters) on top of a picture, such as a scan with a watermark, is OCR'd rather than trusted.

Data and privacy

OCR happens in the Actor's own container with Tesseract; no file, page image or text leaves your run for an outside OCR service. Nothing is fetched besides the links you list, and the text, word boxes and searchable PDFs are written only to your run's storage. Please OCR only files you are allowed to process.

FAQ

Why is a page empty? OCR found no text on it (a blank or photo-only page). It is listed in pagesBlank and not charged.

What does meanConfidence mean? The average confidence Tesseract gives the words on the OCR pages (0-100). Values above 85 usually mean clean text; below 60, check the scan.

How do I call it from code, an agent or a no-code tool? Start a run through the Apify API with {"urls": [...], "outputFormat": "chunks"} and read the dataset; run-sync-get-dataset-items does both in one request for small batches. LangChain and LlamaIndex have Apify loaders that take the dataset rows directly. Make, n8n and Zapier have Apify modules, and AI agents can find and run the Actor through the Apify MCP server.

Why did I get a wordsUrl instead of words in the row? A dataset row is limited to about 9 MB. When a long document with word boxes would exceed that, the words are saved as a JSON record in the run's key-value store and the row links to it.

Which OCR engine is used? Tesseract (Apache-2.0) with the language data packaged by Debian.

Related Actors by the same author

Actor	What it does
PDF to Markdown	Digital documents to Markdown and RAG chunks with headings and tables. Use it for PDFs that have text; use this Actor for scans.
Website Change Monitor	See what changed on a list of web pages since the last run.
ATS Jobs	Every open job of chosen companies from their Greenhouse, Lever, Ashby and other job boards.
RemoteOK & Jobicy Remote Jobs API	Remote jobs from two official APIs, with salary and source credit.

For AI Agents

An agent that meets a scan, a photo of a receipt or a PDF without a text layer can hand it to this Actor and get text back. There is no login and no API key of a third-party OCR service; the run is paid per page through Apify (pay per event) and uses limited permissions.

Connect through the Apify MCP server and expose just this Actor:

{
  "mcpServers": {
    "pdf-ocr": {
      "url": "https://mcp.apify.com?tools=tamatama/image-pdf-ocr"
    }
  }
}

Sign in with OAuth when the client asks, or add the header Authorization: Bearer <APIFY_TOKEN>. The Actor then shows up as its own tool, tamatama--image-pdf-ocr, and the server adds get-dataset-items for paging through long results. Input an agent usually sends:

{ "urls": [{ "url": "https://example.com/scan.pdf" }], "languages": ["eng", "deu"], "outputFormat": "text" }

Tips for agents: in the default text format each file row has meanConfidence (check it before trusting the text) and pagesFailed / error (read them when a row looks empty). Chunk rows (outputFormat: "chunks") carry only the text and page range; page counts and pagesFailed per file are then in the SUMMARY record of the run's key-value store (meanConfidence is only in text rows). Set searchablePdf: true when the user wants a file back; the row then has a searchablePdfUrl.

OpenAPI for tool importers: https://api.apify.com/v2/acts/tamatama~image-pdf-ocr/builds/default/openapi.json. One-request REST call: POST https://api.apify.com/v2/acts/tamatama~image-pdf-ocr/run-sync-get-dataset-items with the JSON input.