PDF to Text, Markdown & Tables (OCR, RAG)

This Actor converts PDF files to plain text and Markdown. It finds headings, paragraphs, bullet lists and simple tables. Scanned pages are read with Tesseract OCR, including Arabic. You get one record per PDF with metadata such as title and page count. You can also get chunks sized for embeddings, ready for RAG and vector databases.

What it does

• Reads PDFs from URLs you give, or from your own Apify key-value stores.
• Returns plain text and Markdown with #, ## and ### headings. Heading levels come from font sizes.
• Finds simple tables and returns them as rows and as Markdown tables.
• Drops running headers and footers, such as page numbers, that repeat on many pages.
• Reads scanned pages with Tesseract OCR. English and Arabic ship with the Actor. 13 more languages are available.
• Turns Arabic presentation-form letters back into normal letters, so Arabic text can be searched and embedded.
• Splits the Markdown into chunks with page numbers and the section heading, if you set a chunk size.
• Returns title, author, subject, dates, language, page count and the file's SHA-256.

The Actor runs pdf.js and Tesseract inside the container. Your files are not sent to any outside service.

What you get

Field	Meaning
url	The PDF URL, or kv://store/KEY for a file in a key-value store
status	ok, or failed with the reason in error. Failed files are free
title	Title from the PDF metadata. If there is none, the first top-level heading on page 1
author, subject, keywords, creator, producer, language	PDF metadata. Empty when the file does not set them
creationDate, modificationDate	Dates from the PDF metadata, ISO 8601
pageCount, pagesProcessed, truncated	Pages in the file, pages read, and whether a page limit cut the file short
ocrPages, ocrPagesSkipped	Pages read with OCR, and scanned pages skipped because of a limit
text	Plain text. Table cells are separated by tabs
markdown	Markdown with headings, paragraphs, lists and tables
headings	Level, text and page of each heading
tables	Page, row count, column count and rows of each table
chunks	Only with a chunk size above 0: text, page range and section of each chunk
pages	Only with Include each page separately: text and Markdown per page
charCount, wordCount	Size of the text
fileName, fileSizeBytes, sha256	File details
passwordProtected	true when the PDF needed a password, false when it opened without one

Sample record from a test run on IRS Form W-9 (a US government work in the public domain). Long fields are shortened here:

{
  "url": "https://www.irs.gov/pub/irs-pdf/fw9.pdf",
  "source": "url",
  "status": "ok",
  "error": null,
  "fileName": "fw9.pdf",
  "fileSizeBytes": 140815,
  "title": "Form W-9 (Rev. March 2024)",
  "subject": "Request for Taxpayer Identification Number and Certification",
  "language": "en-US",
  "creationDate": "2024-03-06T13:18:13.000Z",
  "pageCount": 6,
  "pagesProcessed": 6,
  "truncated": false,
  "ocrPages": [],
  "charCount": 37544,
  "wordCount": 6481,
  "headings": [
    { "level": 1, "text": "Request for Taxpayer Identification Number and Certification", "page": 1 },
    { "level": 1, "text": "General Instructions", "page": 1 },
    { "level": 2, "text": "Purpose of Form", "page": 1 }
  ],
  "markdown": "Form W-9 (Rev. March 2024)\n\n# Request for Taxpayer Identification Number and Certification\n\nDepartment of the Treasury ..."
}

A table from IRS Publication 15-T, page 6, as the Actor returned it:

{
  "page": 6,
  "rowCount": 8,
  "columnCount": 2,
  "rows": [["Payroll period", "Add additional"], ["Weekly", "$226.90"], ["Biweekly", "453.80"], ["Semimonthly", "491.70"]]
}

A chunk from IRS Publication 1 with Chunk size 1500:

{
  "index": 1,
  "text": "communications from the IRS, and to speak to a supervisor about inadequate service.\n\n### 3. The Right to Pay No More than the Correct Amount of Tax\n\nTaxpayers have the ri...",
  "charCount": 1356,
  "pageStart": 1,
  "pageEnd": 1,
  "section": "3. The Right to Pay No More than the Correct Amount of Tax"
}

How to use it

1. Paste the PDF links into PDF URLs. Each link must point straight to a PDF file.
2. For files stored in Apify, pick the stores in PDFs in your Apify key-value stores. Add record keys, or leave the keys empty to read every PDF in the stores.
3. For RAG, set Chunk size for embeddings to 1,000 or 2,000.
4. For scanned Arabic documents, add Arabic to OCR languages.
5. Run the Actor. Download the results as JSON, CSV or Excel, or read them through the API.

Input

Field	Default	Meaning
startUrls	IRS Form W-9	PDF links
pdfPassword	none	PDF password. Opens password-protected PDFs. It is tried on every PDF in the run and never saved
keyValueStores, keyValueStoreKeys	none	Your Apify stores and record keys
ocrMode	auto	auto reads pages that have almost no text layer and hold a large image. off skips OCR
ocrLanguages	["eng"]	OCR languages. Each extra language makes OCR slower
extractTables	true	Find simple tables
removeHeadersFooters	true	Drop running headers and footers
includeText, includeMarkdown	true	Turn either output off to save space
includePages	false	Add text and Markdown per page
chunkSize, chunkOverlap	0, 150	Chunk length and overlap in characters. 0 turns chunks off
maxItems	100	Maximum PDFs per run. 0 means no limit
maxPagesPerPdf	500	Pages read per PDF, from page 1
maxPagesPerRun	5,000	Pages read per run
maxOcrPagesPerRun	50	OCR pages per run
maxFileSizeMb	50	Larger files are skipped

Pricing

You pay per event:

Event	Price
Actor start	$0.02 per run
PDF document, first 20 pages included	$0.003 per PDF ($3 per 1,000 PDFs)
Extra page, above 20 pages in a PDF	$0.0002 per page ($0.20 per 1,000 pages)
OCR page	$0.01 per scanned page ($10 per 1,000 pages)

Failed files are free. Pages with a text layer are never charged for OCR. Examples:

• 100 text PDFs of up to 20 pages: $0.02 + $0.30 = $0.32.
• One 71-page text PDF: $0.02 + $0.003 + 51 x $0.0002 = $0.033.
• One 10-page scan: $0.02 + $0.003 + 10 x $0.01 = $0.123.

The Actor never does work it cannot charge for. When your maximum cost per run is close, it reads fewer pages and lists the skipped scanned pages.

Limits and good to know

• Each download is capped at 60 seconds and retried at most twice.
• Only fetch files you have the right to use. The Actor reads robots.txt before each download and skips blocked files. It refuses private network addresses.
• Heading levels come from font sizes. Documents that use bold text of the same size for headings get fewer headings.
• Table detection works for tables whose cells line up in columns. Merged cells and multi-row headers can come back split. Tables on OCR pages come back as text.
• The Actor runs with 1,024 MB of memory. OCR of one dense 300 dpi page took about 35 seconds on the platform at 512 MB. It should take about half that at 1,024 MB, which is not yet measured. Large scans are scaled to at most 2,500 pixels on the long side. OCR stops 2 minutes before the run timeout, so finished work is saved.
• English uses Tesseract's fast model. Arabic and the other languages use the more accurate integer model.
• Password-protected PDFs open when you give the password. Without it, they fail with a clear, free error. Very large results go to the run's key-value store, and the record keeps the first 200,000 characters with a pointer in fullOutputKey.
• Right-to-left text comes back in reading order. Lines that mix Arabic and Latin, such as URLs, can come back partly reordered.

FAQ

Can it read password-protected PDFs? Yes. Put the password in PDF password. The Actor tries it on every PDF in the run. PDFs without a password open as usual. A missing or wrong password gives a failed row with the reason, and failed rows are free. The password never appears in the output or the log.