OCR Citation Chunker

Give this Actor a scanned or image-only PDF and it returns deterministic, page-aware citation-ready JSON chunks with OCR confidence, quality signals, source provenance, and stable citation identifiers.

Use it to turn archival scans, printed reports, forms, and image-based research PDFs into retrieval-ready records for RAG, search, and agent workflows. It is intentionally narrow: it processes PDFs, does not interpret their meaning, and never invents text.

Input

Provide either a single url or a batch in urls. The Actor downloads each PDF, runs Tesseract OCR page by page, and groups usable text into chunks without crossing document boundaries. chunkSize controls the approximate maximum characters per chunk; overlap retains trailing context in the next chunk.

{
  "url": "https://example.com/scanned-report.pdf",
  "chunkSize": 1000,
  "overlap": 100
}

PDF URLs must be publicly reachable. Password-protected, malformed, or unreachable inputs are reported in the final SUMMARY; a failed document does not discard usable output from other documents in the same run.

Output

The default dataset contains one JSON record per citation-ready chunk. Important fields include:

• text, pageStart, and pageEnd for grounding a citation.
• documentId (SHA-256 of source bytes) and deterministic chunkId for idempotent downstream indexing.
• sourceUrl, resolvedUrl, sourceByteSize, and documentSha256 for provenance.
• ocrConfidence plus ocrQualityScore, ocrQualityTier (high, usable, or low), and ocrQualityFlags for filtering uncertain extraction.
• characterCount and estimatedTokens to help size vector-store or LLM work.

The key-value store record SUMMARY reports totals, per-document outcomes, failed URLs, and average OCR quality. A useful OCR page is one that produces usable text and saved citation-ready output. Blank or non-readable pages are not emitted as chunks, are not charged as useful OCR pages, and are reflected in document outcomes.

Example record

{
  "documentId": "sha256:...",
  "chunkId": "sha256:...",
  "pageStart": 4,
  "pageEnd": 4,
  "ocrConfidence": 92.4,
  "ocrQualityScore": 94.1,
  "ocrQualityTier": "high",
  "ocrQualityFlags": [],
  "text": "Extracted source text..."
}

Limits and quality

OCR is best for clear, reasonably upright scans. Handwriting, complex tables, faint pages, unusual scripts, and heavily compressed imagery may produce incomplete text. Use ocrQualityTier and ocrQualityFlags as deterministic review signals before relying on a chunk. The Actor currently targets English OCR.

The production configuration uses 512 MB with OMP_THREAD_LIMIT=1. A clean 50-page cloud control completed successfully with stable late-page timing. Actual cost varies with page count, image resolution, and source quality. See BENCHMARKS.md in the source for measured fixtures and run evidence.

Pricing

The Actor uses pay-per-event pricing. It charges once for each useful OCR page only after that page's citation-ready output is saved. Blank pages that produce no useful output are not charged as page-processed events. Platform usage costs are included.

Subscription tier	Useful OCR page price
Free / Bronze	$1.50 per 1,000 pages
Silver	$1.35 per 1,000 pages
Gold and higher	$1.20 per 1,000 pages

Actor start is $0.00005 per run.

Integration

Consume the dataset through the Apify API, webhook, Make, Zapier, or an agent workflow. Store chunkId to deduplicate re-runs, cite sourceUrl with pageStart/pageEnd, and retain documentId to trace every chunk back to its exact source bytes.

For AI agents and MCP

Contract: public scanned or image-only PDF URL → OCR-backed, page-aware citation-ready JSON chunks.

Use this Actor for scanned PDF OCR, image-only PDFs, OCR to JSON, citation-ready OCR, and scanned-document RAG ingestion. For PDFs with an embedded text layer, use PDF Citation Chunker instead.