PDF to Markdown + OCR — Structured Document Extraction

Public PDF URL in → structured Markdown out. Born-digital pages extract natively; scanned / image-only pages go through a managed OCR engine. Built for PDF to Markdown, PDF OCR API, scanned PDF extraction, and structured document data pipelines (MCP / API / schedule).

Actor id	kaz_kakyo/pdf-structured-ocr
Minimal input	{ "pdfUrls": ["https://…/file.pdf"] }
Cost	$0.001/native page · $0.012/OCR page · $0.005/document manifest · $0.00005 start
Output	type: "page" rows + one type: "manifest" per document (or type: "error")

Call it (MCP / API)

{
  "actor": "kaz_kakyo/pdf-structured-ocr",
  "input": {
    "pdfUrls": ["https://example.com/report.pdf"],
    "ocr": "auto"
  }
}

OCR is included — no setup, no key required. Native-text pages always succeed; if a page needs OCR and OCR capacity is unavailable, it is marked mode: "unavailable" and listed in partialFailures. Very large or image-heavy pages may return unavailable rather than being silently truncated. Billed only for pages actually delivered.

Because OCR capacity is shared across all users, each run caps OCR-billable pages (maxOcrPages, default 100, hard ceiling 200). Native (born-digital) extraction is unaffected by the cap and consumes no OCR capacity.

import { ApifyClient } from 'apify-client';

const client = new ApifyClient({ token: process.env.APIFY_TOKEN });
const run = await client.actor('kaz_kakyo/pdf-structured-ocr').call(
  {
    pdfUrls: ['https://example.com/report.pdf'],
    ocr: 'auto',
    ocrMaxWaitSecs: 300,
    maxPages: 50,
  },
  { maxTotalChargeUsd: 1.0 },
);
const { items } = await client.dataset(run.defaultDatasetId).listItems();
const manifest = items.find((i) => i.type === 'manifest');

Output contract

Page row (charged only when mode is native or ocr):

{
  "type": "page",
  "url": "https://example.com/report.pdf",
  "pageNumber": 1,
  "mode": "native",
  "confidence": null,
  "markdown": "## Introduction\n\n…",
  "textCharCount": 1842
}

mode	Meaning
native	Born-digital text extracted locally
ocr	Page recovered by the managed OCR engine
unavailable	Needed OCR but capacity unavailable / OCR failed / budget exhausted — never charged

Manifest (charged only when ≥1 page was delivered):

{
  "type": "manifest",
  "url": "https://example.com/report.pdf",
  "pageCount": 12,
  "pagesDelivered": 12,
  "markdown": "<!-- page 1 (native) -->\n\n…",
  "pages": [{ "pageNumber": 1, "mode": "native", "confidence": null }],
  "partialFailures": []
}

Pricing

Event	Price	When
native-page-delivered	$0.001	Each saved native-text page
ocr-page-delivered	$0.012	Each saved OCR page
document-manifest-delivered	$0.005	Final manifest, only if ≥1 page delivered
apify-actor-start	$0.00005	Per run
Default dataset item	$0	Never billed separately

Charge order: budget-check → save dataset row → Actor.charge(). Unavailable / error rows are never billed. Cap spend with maxTotalChargeUsd.

Input

Field	Notes
pdfUrls / pdfUrl	Public http(s) PDF links only
maxPages	Default 50 (hard cap 500)
ocr	auto (default) · never · always
maxOcrPages	Default 100 (hard ceiling 200); per-run OCR page cap
ocrMaxWaitSecs	Default 300 (max 900); OCR timeout → ocr_timeout, uncharged

Private / localhost / link-local / RFC1918 targets are rejected (SSRF guard, enforced at connect time and on every redirect hop). File size cap: 50 MB, enforced while streaming. PDFs are fetched over direct egress only — no proxy, because a proxy cannot enforce the connect-time address ACL. OCR processes the already-fetched PDF bytes; the source is never re-fetched.

Why this one

• Structured Markdown, not a blob of OCR text — headings and crude tables from layout when native text exists.
• Page-level honesty: every page reports mode + confidence; gaps go in partialFailures.
• OCR only when needed (auto) so born-digital PDFs stay cheap.
• Billed only for pages actually delivered — timeouts, failures, and capacity limits never charge OCR pages.