Document Text Extractor: PDF to Text, Markdown and OCR

Turn PDFs, Word documents and scanned images into clean text and markdown. Give the Actor a list of document URLs or uploaded files and it returns the text of each one as plain text and as markdown, with per-page detail and metadata. Pages with no text layer (scans) are read with OCR. It is built for RAG and LLM pipelines, document automation and search indexing. Pricing is pay per result: $0.006 per document and $0.010 per OCR'd page. No start fee, and failed documents are free.

Run it from the Apify Console or call it through the Apify API.

Use cases

• RAG and LLM ingestion. Convert PDFs to markdown, one record per page, ready to chunk and embed.
• Document automation. Pull the text, title, author and page count out of incoming PDFs and Word files.
• Search indexing. Bulk-extract text from up to 100 documents per run to feed a search index.
• Digitising scans. Read scanned PDFs and photographed pages with English OCR.
• Research papers. Extract text from arXiv-style PDF links in bulk.
• Metadata audits. Collect title, author, producer, created date and page count across a set of PDFs.

Supported formats

• PDF: text-layer PDFs, scanned PDFs (OCR), and mixed ones. PDFs with only permission restrictions ("no printing") are extracted normally.
• DOCX (Word): text and markdown, using the document's own heading styles.
• Images: PNG, JPEG and TIFF, read with OCR as a single page.

The format is detected from the file's contents, not its name or Content-Type. Anything else returns an unsupported format error, free of charge.

PDF to text

Pass PDF URLs (or uploaded files) and get the plain text of each document, along with the page count, title, author and other metadata. This works for any link that serves the file directly, including arXiv PDF links. Set format to text if you only want plain text.

PDF to markdown for RAG and LLMs

Set format to markdown (or leave it on both) to get markdown with headings and paragraphs. For RAG, set outputMode to page: you get one dataset item per page, which is a natural unit for chunking and embedding. For a quick prompt, document mode returns the whole document in one item.

Markdown heading detection is best effort. In PDFs, headings come from font size relative to the body text: the largest size becomes # and smaller heading sizes become ##. Paragraphs come from line gaps. DOCX files use their own heading styles. Markdown problems never cause an error. Tables and layout are not reconstructed.

OCR for scanned PDFs and images

Pages with no text layer are read with OCR, and image files (PNG, JPEG, TIFF) are OCR'd as a single page.

• English only. Other languages may come out garbled.
• auto (default) OCRs only pages whose text layer has fewer than 50 characters, which is usually scanned pages.
• force OCRs every page, up to maxOcrPagesPerDocument. Use it when a PDF's text layer is broken or garbled.
• off never OCRs. Images then fail with image requires OCR, free of charge.
• Once a document reaches maxOcrPagesPerDocument (default 20, up to 50), OCR stops. The remaining pages keep their text layer, which may be empty.
• Each OCR page has a 20-second limit. A page that times out keeps its text layer and is not billed.

Each page in the output carries an ocr flag, so you can see which pages were OCR'd.

DOCX to text and markdown

Word documents are converted using the document's own heading styles, so headings in Word become headings in the markdown. Use format to choose text, markdown or both. Word documents return no title or author metadata, and they count as one page.

Per-page output and metadata

• Per page. outputMode: "page" gives one item per page. In document mode, every page is inside the item's pages array.
• Metadata. Each item includes title, author, producer, created, pageCount, fileSizeBytes, charCount, ocrPages and truncated. Values the Actor cannot determine are null.

Pricing

Pay per result, with no start fee:

• $0.006 per document successfully extracted (document event).
• $0.010 per OCR'd page (ocr-page event), on top of the document charge.

Worked examples:

Job	Cost
100 text-layer PDFs, no OCR	100 x $0.006 = $0.60
One 10-page scanned PDF, all pages OCR'd	$0.006 + 10 x $0.010 = $0.106
100 DOCX files	100 x $0.006 = $0.60
One PNG image (OCR'd as one page)	$0.006 + $0.010 = $0.016

Failed documents are free: a download error, an unsupported or password-protected file, a file over the size cap, or a document that times out produces an error record and costs nothing. OCR pages that fail or time out are not charged. In page mode you still pay per document, not per page item. The SUMMARY record is not billed.

Charge limit. If your run hits its charge limit, the Actor stops before downloading the next document and marks the rest skipped in SUMMARY.

Real-world benchmark

From a real run on 2026-10-21, OCR off (see docs/samples/federal-register-run-2026-10-21.json):

• 60 Federal Register PDFs (public-domain US government documents).
• 60 of 60 extracted: 151 pages in 10 seconds.
• Text per document ranged from about 5,000 to 116,000 characters (median about 13,700).
• User cost: $0.36 at $0.006 per document.

Excerpt from the first document (https://www.govinfo.gov/content/pkg/FR-2026-10-21/pdf/C1-2026-20311.pdf):

66591 Federal Register / Vol. 91, No. 202 / Wednesday, October 21, 2026 / Notices
1 15 U.S.C. 78s(b)(1).
2 15 U.S.C. 78a.
3 17 CFR 240.19b–4.
4 See Securities Exchange Act Release No. 106204
(Aug. 27, 2026), 91 FR 56247 (Sept. 1, 2026)

Input

{
  "documents": [
    "https://arxiv.org/pdf/1706.03762"
  ],
  "outputMode": "document",
  "format": "both",
  "ocr": "auto",
  "maxOcrPagesPerDocument": 20,
  "maxFileSizeMb": 25
}

• documents (required): 1 to 100 items. Each is an http(s) URL or a file you upload in the input form.
• outputMode: document (default) gives one item per document with every page inside it; page gives one item per page, which suits RAG pipelines.
• format: both (default), text or markdown.
• ocr (advanced): auto (default), force or off. See OCR above.
• maxOcrPagesPerDocument (advanced): 0 to 50, default 20.
• maxFileSizeMb (advanced): 1 to 50, default 25.

Output

One dataset item per document (document mode):

{
  "sourceUrl": "https://example.com/quarterly-report.pdf",
  "fileName": "quarterly-report.pdf",
  "mimeType": "application/pdf",
  "status": "ok",
  "error": null,
  "text": "Quarterly Report\nRevenue grew in every region.\n\nPage two has more body text.",
  "markdown": "# Quarterly Report\n\nRevenue grew in every region.\n\nPage two has more body text.",
  "pages": [
    { "page": 1, "text": "Quarterly Report\nRevenue grew in every region.", "ocr": false },
    { "page": 2, "text": "Page two has more body text.", "ocr": false }
  ],
  "metadata": {
    "title": "Quarterly Report",
    "author": null,
    "producer": "Example Writer",
    "created": "2026-09-30T12:00:00.000Z",
    "pageCount": 2,
    "fileSizeBytes": 18234,
    "charCount": 75,
    "ocrPages": 0,
    "truncated": false
  }
}

In page mode each item holds one page: pages is replaced by top-level page and ocr, and the document fields repeat. text is left out when format is markdown, and markdown when it is text. Values the Actor cannot determine are null.

A failed document gives exactly one item with "status": "error" and the reason in error, such as HTTP 404, unsupported format, encrypted, could not parse or file too large (> 25 MB cap). The run carries on with the next document.

The Actor also writes a SUMMARY record to the run's key-value store, with one entry per input (sourceUrl, status, pageCount, ocrPages and any error), plus totals and chargeLimitReached.

If one item would be larger than about 4 MB, its text is shortened, the full JSON is saved to the run's default key-value store, and fullOutputUrl points to it.

Limits

• Up to 100 documents per run, processed 2 at a time.
• Files up to 25 MB by default, 50 MB at most.
• Text extraction reads the first 500 pages of a document. metadata.truncated is true when there are more.
• Up to 50 OCR pages per document (default 20).
• Each document has a 10-minute budget. A document that runs over it becomes a free error record.
• Downloads time out after 60 seconds.
• Word documents return no title or author metadata, and they count as one page.

FAQ

Which file formats are supported? PDF (text-layer, scanned and mixed), DOCX, and PNG, JPEG and TIFF images. The format is detected from the file's contents. HTML, XLSX and PPTX files are not supported and return a free unsupported format error.

What are the limits? 100 documents per run, 25 MB per file by default (up to 50 MB with maxFileSizeMb), 500 pages read per document, and up to 50 OCR pages per document.

Does OCR work in languages other than English? No. OCR is English only, and other languages may come out garbled.

Can it read password-protected PDFs? No. A PDF that needs a password to open returns an encrypted error and is not charged. PDFs that open without a password but have permission restrictions, such as "no printing", are extracted normally.

Are my files stored? No. Files are not stored after the run. See Privacy below.

Should I use URLs or uploads? Either. A URL must be an http(s) link that serves the file. Otherwise, upload files in the input form's documents field and mix them with URLs as you like, up to 100 items in total.

Does it extract tables or keep the page layout? No. Tables and layout are not reconstructed. You get text, plus markdown headings and paragraphs detected on a best-effort basis.

Privacy

Files are not stored after the run. Documents are held only in memory and in a temporary folder that is deleted as soon as each document is done. Uploaded files are read with your run's own token, which is never sent to any other host. The only things saved are the dataset items, SUMMARY, and any oversized record, all in the run's own default storages.

Out of scope

Tables and layout reconstruction, non-English OCR, password-protected PDFs, HTML, XLSX and PPTX files, AI summaries, and long-term storage.