Scanned PDF/Image OCR to Markdown

Turn scanned PDFs and images into Markdown with RapidOCR (onnxruntime) — CJK-capable, no external AI API keys.

This Actor rasterizes each PDF page (or loads an image), runs offline OCR, and writes Markdown with optional page markers and RAG chunks. Field names for chunking mirror the digital PDF/DOCX Actor where practical.

What you get

• OCR for scanned PDFs and common images (PNG, JPEG, WebP, TIFF, …)
• Default engine: RapidOCR + onnxruntime (Apache-2.0 / MIT stack; see NOTICE)
• Optional page markers and RAG chunks (chunkSize / chunkOverlap / chunkUnit)
• Best-effort table assembly from OCR box geometry — not the digital PDF word-index table repair used by the PDF/DOCX Actor
• Dataset items + optional .md key-value store files

Measured cloud runs (2026-09-30, Asia/Taipei)

Own runs on this Actor, build 0.1.3, memory 2048 MB. Platform cost = usageTotalUsd re-read after finish (not PPE revenue). Public-domain LOC/WDL scans only.

Run	Input	Pages	Wall	Peak RSS	CU	Platform USD
huicge2RO7IJnpaQs	Emancipation Proclamation (image PDF)	5	109 s	1077 MB	0.061	$0.0123
OdyOeoSfhaaghyr9t	Furness sermon 1841, pages 1–8	8	42 s	1361 MB	0.023	$0.0048
e5xfmZVXzIZwjQexI	RapidOCR CJK+EN sample JPG	1	9.5 s	310 MB	0.005	$0.0012

Dense-page platform cost ≈ $0.0025 / page on the Emancipation scan (large page images @ 200 DPI). Sparse cover pages are cheaper. These runs confirm OCR completes and writes dataset items; they are not an accuracy benchmark. Old print and manuscript-style pages show typical OCR noise (misread characters). Do not expect digital-PDF table quality.

Use cases

• Ingest scanned reports, pamphlets, and image-only PDFs into Markdown for RAG or note-taking
• Batch OCR of screenshots / phone photos of documents (print, not handwriting)
• CJK + Latin mixed scans where a local RapidOCR stack is enough (no cloud vision API)

How to use

1. Paste direct URLs to scanned PDFs/images, or upload files.
2. Choose markdown, chunks, or both.
3. Optionally set pageRange, dpi (PDF), and maxPagesPerDocument.

Example input

{
  "urls": [{ "url": "https://example.com/scan.pdf" }],
  "outputFormat": "markdown",
  "pageMarkers": "comment",
  "dpi": 200,
  "maxPagesPerDocument": 10
}

Example output (dataset document item, fields abbreviated)

{
  "type": "document",
  "status": "success",
  "fileName": "scan.pdf",
  "pageCount": 5,
  "pagesConverted": 5,
  "stats": { "ocrItems": 105, "tables": 0, "words": 728, "dpi": 200 },
  "markdown": "<!-- page: 1 -->\n\n…",
  "markdownKey": "scan.md"
}

Pricing

Pay-per-event:

Event	Price
Actor start	Apify default ($0.00005 / GB memory)
Document OCR'd	$0.005 per successful document
Page OCR'd (primary)	$0.006 per page

Failed downloads / unreadable files are never charged. RAG chunking does not add events.

Worked example: 1 document × 10 pages → $0.005 + 10 × $0.006 = $0.065 (plus the synthetic start event).

Known limits

• Handwriting is unreliable; this Actor targets print scans.
• Japanese / Korean: RapidOCR can return some text, but this release was not accuracy-tested on JP/KR corpora — treat as best-effort.
• Complex multi-column layouts may read top-to-bottom incorrectly.
• Table assembly ≠ digital repair: geometry from OCR boxes only; do not expect the PDF/DOCX Actor's word-index table quality.
• Very large / high-DPI PDFs need more memory and time (default run memory 2048 MB; multipage LOC peaks observed ~1.1–1.4 GB RSS).
• Prefer the digital PDF/DOCX Actor when a text layer already exists.

License & source code

This Actor is open source under the GNU Affero General Public License v3.0 (AGPL-3.0) — see LICENSE. The full source code is public: https://github.com/xbox002000/scanned-ocr-to-markdown

Third-party notices: NOTICE.