PDF Text & Table Extractor — Markdown, OCR

Unofficial tool. Not affiliated with, endorsed by or sponsored by Google (PDFium) or the maintainers of Tesseract. Built with PDFium (BSD-3) and Tesseract (Apache-2.0); all trademarks belong to their owners. Removal requests: TAKEDOWN.md — honoured in 48 hours. Privacy: PRIVACY.md.

Paste PDF URLs, get text back — as one row per page, one row per document, or as RAG chunks with page and character offsets. PDFium does the reading: a 142-page US tax publication came back in 1.2 s in our measurements. Optional heuristic markdown, ruled-table extraction and Tesseract OCR for scanned pages ride along in the same Actor. You pay per page delivered; the document summary row, error rows, duplicates and pages you skipped cost nothing.

What it extracts

Content	Delivered?	Notes
Text layer	Yes	PDFium text extraction, \n line breaks, Unicode NFC
Markdown	Yes	Headings from font-size rank, paragraphs from line gaps — a heuristic, not a layout model
Ruled tables	Yes	Tables drawn with ruling lines, as arrays of rows and as GitHub tables in the markdown
Borderless tables	No	Borderless tables stay as text
OCR (English)	Opt-in	Tesseract on image-only pages, charged as ocr-page
Images	No	Not exported
Form fields	No	Field values are not read
LLM post-processing	No	Deterministic: identical bytes and input give identical charged rows

How to use

1. Paste direct PDF links into urls, one per line — the file itself, not a viewer page.
2. Pick an output mode. page for spreadsheets and Google Sheets, document for one item per file, chunk for a vector store. Turn on extractTables or ocr if you need them.
3. Run it. Set maxPages (default 5,000) or the run's maximum charge to bound the spend. The input form shows every option with its default; the prefill runs the six-page IRS Form W-9.

Output modes

Page (recordType: "page", charged) — one row per page. The example is page 1 of the IRS Form W-9 (https://www.irs.gov/pub/irs-pdf/fw9.pdf, a US-government public-domain work, 6 pages, 141 KB) exactly as the prefill run produced it; only text and markdown are cut short:

{
  "recordType": "page",
  "url": "https://www.irs.gov/pub/irs-pdf/fw9.pdf",
  "documentId": "2d420cbb4123dcf1",
  "page": 1,
  "pageCount": 6,
  "text": "Form W-9\n(Rev. March 2024)\nRequest for Taxpayer \nIdentification Number and Certification\nDepartment of the Treasury \nInternal Revenue Service ...",
  "markdown": "# Form W-9\n\n(Rev. March 2024)\n\n## Request for Taxpayer\n\n## Identification Number and Certification\n\nDepartment of the Treasury Internal Revenue Service ...",
  "tables": null,
  "tableCount": 0,
  "charCount": 5704,
  "wordCount": 933,
  "width": 611.98,
  "height": 791.97,
  "rotation": 0,
  "hasTextLayer": true,
  "ocrApplied": false,
  "needsOcr": false
}

Document (recordType: "document") — one row per URL. It is always written, free, as the last row for that document, and carries status, errorCode, documentId, fileName, HTTP status, byte size, page counts, PDF metadata, encrypted, permissionsCopyAllowed, durationMs and fetchedAt. In outputMode: document the same row also carries text, markdown, pages[{page, text, charCount, ocrApplied}], tables, charCount and wordCount, and is charged by page count after it has been pushed. A viewer page instead of a PDF comes back like this:

{
  "recordType": "document",
  "url": "https://docs.google.com/viewer?url=...",
  "status": "error",
  "errorCode": "not_pdf",
  "httpStatus": 200,
  "contentType": "text/html; charset=utf-8",
  "bytes": 0,
  "pageCount": null,
  "pagesExtracted": 0,
  "pagesCharged": 0
}

Chunk (recordType: "chunk") — chunkSize characters (default 1,500 ≈ 375 tokens), split at paragraph, then sentence, then hard cut, with chunkOverlap repeated from the previous tail. Each row has chunkIndex, chunkCount, pageStart, pageEnd, charStart, charEnd, tokenEstimate and headingPath (the nearest markdown headings above the chunk). Chunks are charged per page covered, so the chunk size never changes the bill.

Error codes on document rows: blocked_url, download_failed, too_large, not_pdf, password_required, permissions_restricted, malformed, timeout, pages_capped, budget_exhausted, run_timeout, item_too_large, duplicate. Two dataset views are provided: Pages and Documents.

Coming from memo23, automation-lab or gochujang

Their input	Here
urls	urls (same)
pageRange (memo23)	pageRange (same syntax: 1-5, 8, 12-)
startPage / endPage	pageRange
ocrScannedPages / enableOcr	ocr
maxPdfSizeMb / maxPdfMb	maxPdfMb (same)

What you gain: per-page billing instead of per-file, ruled tables, RAG chunks with offsets, a documentId for deduplication, typed error codes on free rows, and hasTextLayer / needsOcr flags so you know before paying for OCR.

What you lose, stated plainly: no file upload or base64 input (URLs only), no proxy retry on a refused download, English OCR only, and flat-price Actors are cheaper for long text documents — see Pricing.

Pricing

$0.0003 per page, $0.003 per OCR page, no start fee. Free, always: document summary rows, error rows, duplicate rows, pages outside pageRange, pages beyond maxPagesPerPdf or maxPages, image-only pages that were not OCRed (ocr off, maxOcrPagesPerPdf reached or Tesseract failed — delivered empty with needsOcr: true), and documents never reached because the run stopped on budget or before its timeout.

Worked examples, plain arithmetic:

Job	Pages	Cost
2-page invoice	2 × $0.0003	$0.0006
10-page paper	10 × $0.0003	$0.003
100-page manual	100 × $0.0003	$0.03
300-page manual (maxPagesPerPdf: 300)	300 × $0.0003	$0.09
20 scanned pages with OCR	20 × $0.003	$0.06

Comparable Actors on the Store, public prices only, as of 2026-08-28:

Actor	Model	2 pp	10 pp	100 pp	20 pp OCR
this Actor	$0.0003/page, $0.003/OCR page, no start fee	$0.0006	$0.003	$0.03	$0.06
memo23 pdf-text-extractor	$0.005 start + $0.005/PDF + $0.015/OCR page	$0.01	$0.01	$0.01	$0.31
gochujang pdf-text-extractor	$0.001 start + $0.02/PDF + $0.0005/page	$0.022	$0.026	$0.071	$0.031 (no OCR event)
automation-lab pdf-text-extractor	price not public	—	—	—	—

Per-page pricing wins on short documents, on runs where many downloads fail or are skipped, and on OCR. For text documents longer than about 17 pages memo23's flat price is lower. Set maxPagesPerPdf (default 100) and maxPages to keep long documents within the budget you intend; when the run's maximum charge is reached the Actor stops pushing, writes free budget_exhausted rows for what it did not reach, and still finishes successfully.

Security and limits

• Private and internal URLs are refused before any request (blocked_url): only http/https on ports 80, 443, 8080 and 8443, no localhost, no .internal / .local, no private, loopback, link-local, multicast or reserved addresses. The resolved address is pinned for the request, so DNS cannot be swapped between the check and the download; redirects are re-checked on every hop, at most five, never from https to http. Hostnames listed in the Actor's blocklist.txt (removal requests, see TAKEDOWN.md) are refused the same way, before any request.
• Files over maxPdfMb (default 50 MB, maximum 100 MB) are refused on Content-Length and aborted during download. The effective cap is also bounded by the run's memory (memory / 4). Each download hop has a 70-second wall (10 s connect + 60 s body); a server that stalls or trickles bytes gets a free timeout row instead of holding the run.
• Each document is parsed in its own memory-capped subprocess with a wall-clock limit (120 s, 600 s with OCR), so a hostile or broken file becomes one free malformed or timeout row and the run continues.
• One plain GET per URL, one download in flight per host and at least one second between requests to the same host, no retry on 401, 403 or 429, no proxy, no cookies (a Set-Cookie is dropped, never replayed), no connection reuse between hosts, no login.
• No DRM or paywall bypass. Owner-restricted (no-copy) PDFs are skipped with permissions_restricted; there is no override. User-password-protected files need pdfPassword.
• 100 URLs per run, 100 pages per PDF by default (up to 2,000), 2 documents in flight.

Privacy

The output contains whatever your documents contain. If a document holds personal data, you are the controller of the resulting dataset. The Actor opens no outbound connection to the developer and embeds no developer credential, so the Actor T&C creator-access clause is not exercised; your URL list is visible to us only through Apify's Console if you open an issue about a run. The file exists only inside the run container (memory or its ephemeral disk) and is gone when the run ends; the dataset lives in your Apify account under Apify's retention — delete the run to erase it. We do not log URLs or document text.

For AI agents and MCP

Use outputMode: "document" for one call → one item with the full text, or "chunk" to feed a vector store directly. maxPages and the run's maximum charge are honoured before extraction, so an agent never pays for pages it cannot receive, and with ocr off there is a single price to reason about. Errors are typed rows, not exceptions.

{ "urls": ["https://www.irs.gov/pub/irs-pdf/fw9.pdf"], "outputMode": "document", "maxPages": 50 }

Limitations you should know before you buy

• Ruled tables only; borderless tables stay as text. Table detection adds about 0.3–0.5 s per page.
• Multi-column reading order follows the PDF's content stream. PDFium kept column order on a two-column paper in our tests, but there is no guarantee.
• Headings are a font-size heuristic; a document with a single font size gets no headings.
• OCR measured 93–97 % similarity on clean 150–200 dpi scans of printed English; worse on handwriting, skew or low resolution. English only.
• Encrypted files need pdfPassword; owner-restricted files are always skipped.
• No file upload; URLs only. 100 URLs per run; 100 pages per PDF by default (up to 2,000); one request per host per second.
• The run-timeout guard stops before the platform timeout and writes free run_timeout rows, so a run never ends as TIMED-OUT. At defaults that is roughly 100 text PDFs or 20 OCR PDFs per run.
• A document-mode row over 8 MB drops markdown and pages[].text (

  errorCode: item_too_large
  ); text is kept, so the page charge stands. If text alone still exceeds the platform's 9 MB item limit, the row is delivered as a free item_too_large error row without content and nothing is charged — use page or chunk mode for such documents.
• Image size: 628 MB as measured on the first build (PDFium, pdfplumber, Tesseract and the English language pack).

FAQ

Is this legal? The Actor sends one plain GET per URL you supply, with an identifying User-Agent, no login, no proxy and no retry on a refusal. The text belongs to the document's copyright owner: you must have the right to download and process each file, and you must follow the hosting site's terms and robots rules. Owner-restricted PDFs are skipped. Rights-holders who want a file excluded from future runs are pointed at the hosting site or the user who supplied the URL — see TAKEDOWN.md.

A scanned PDF returns empty text? Turn on ocr. Image-only pages are delivered empty and free with needsOcr: true while it is off.

How do I know what I will pay before paying for OCR? Run once with ocr off: every page row carries hasTextLayer and needsOcr, and the document row counts imageOnlyPages.

Google Drive link? Use the direct-download form of the link. A viewer page is HTML and comes back as a free not_pdf row.

How do I dedupe? documentId is a hash of the file bytes: the same file behind two URLs gets one set of charged rows and a free duplicate row.

What is pageRange syntax? 1-5, 8, 12- — 1-based, 12- means 12 to the end. Pages outside the range are never charged.

Support

Open an issue on the Actor's Issues tab or in the public repository at https://github.com/moonie0201/apify-utils. Every issue gets a reply within 14 days, usually within 48 hours. Bug reports that include the run id and the input are fixed fastest. Removal, takedown, copyright or privacy requests jump the queue: mooniegilog@gmail.com · TAKEDOWN.md · PRIVACY.md.

Disclaimer

You are responsible for having the right to download and process each file and for the hosting site's terms and robots rules. The Actor sends one plain GET per URL with an identifying User-Agent, no login, no proxy and no retry on a refusal; it bypasses no DRM, paywall or owner restriction. This tool is unofficial and is not affiliated with, endorsed by or sponsored by Google (PDFium), the Tesseract maintainers or any site whose files you process. All trademarks belong to their respective owners.