PDF to Markdown, Tables & RAG Chunks (with OCR)

Turn PDF, Word (DOCX) and image files into clean Markdown, tables (rows + CSV) and RAG-ready chunks with page numbers. Scanned pages are read automatically with OCR. No API keys, no LLM calls, nothing to configure. You pay a fraction of a cent per page.

Built for AI agents and LLM pipelines: give it links (or upload a file), get back text your model can actually use.

What you get for each document

Field	What it is
markdown	The whole document as Markdown: headings (#, ##), paragraphs, bullet lists, tables, and <!-- page N --> markers
tables	Every detected table with its page, rows (list of lists) and csv
chunks	Text split for embeddings: text, page_start, page_end, nearest heading, char_count
metadata	Title, author, subject, dates from the file properties
page_count, pages_text, pages_ocr, pages_empty	What was read and how
warnings, error	Plain-language notes and stable error codes (see below)

Each document's Markdown is also saved as a .md file in the run's key-value store.

Quick start

In the Console: paste one or more links into Document URLs (or upload a file) and click Start.

API (Python):

from apify_client import ApifyClient

client = ApifyClient("<YOUR_APIFY_TOKEN>")
run = client.actor("errandkit/pdf-to-markdown-tables").call(run_input={
    "documentUrls": ["https://www.irs.gov/pub/irs-pdf/fw9.pdf"],
    "outputMode": "documents",
})
for item in client.dataset(run["defaultDatasetId"]).iterate_items():
    print(item["document_name"], item["page_count"], len(item["tables"]))
    print(item["markdown"][:500])

RAG ingestion (one dataset item per chunk, ready for a vector database):

{
  "documentUrls": ["https://example.com/handbook.pdf", "https://example.com/policy.docx"],
  "outputMode": "chunks",
  "chunkMaxChars": 1200,
  "chunkOverlapChars": 150,
  "includeMarkdown": false
}

Base64 input (for agents and back ends that already hold the file):

{"documents": [{"name": "invoice.pdf", "content_base64": "JVBERi0xLjcK..."}]}

Use it from an AI agent (MCP)

This Actor works as a tool in Claude, Cursor, ChatGPT connectors and other MCP clients through Apify's MCP server:

https://mcp.apify.com/?actors=errandkit/pdf-to-markdown-tables

The agent passes a link, the Actor returns Markdown, tables and chunks. Results are JSON with stable field names, so agents don't have to guess.

Input options

Option	Default	Notes
documentUrls	–	Direct links to PDF, DOCX, PNG, JPEG, TIFF, GIF, BMP or WEBP files
uploadedFile	–	Upload one file in the Console
documents	–	JSON list of {name, url} or {name, content_base64}
outputMode	documents	documents = one item per file; chunks = one item per chunk
pages	all	PDFs only, e.g. 1-3,7
ocr	auto	auto reads only pages without a text layer; force OCRs every page; off never
ocrLanguage	eng	spa, fra, deu, por, ita, or combined like eng+spa
tables	true	Detect ruled tables in PDFs; all DOCX tables are kept
chunking, chunkMaxChars, chunkOverlapChars	true, 1500, 150	Chunks keep the nearest heading and page range; tables are split by rows with the header repeated
includeMarkdown	true	Turn off to keep items small
saveMarkdownFiles	true	Save 001-name.md files to the key-value store
maxPagesPerDocument	500	Pages after this are skipped and not charged
failOnEmpty	true	Mark the run failed if no document has any text

Pricing

Pay per event. You only pay for pages that produced text.

Event	Price	Per 1,000
Page processed (text layer; DOCX counts 1 page per 3,000 characters)	$0.0005	$0.50
Page or image read with OCR	$0.003	$3.00
Actor start (Apify standard)	$0.00005	$0.05

Examples: a 10-page digital PDF costs $0.005; a 10-page scan costs $0.03. Blank pages, failed downloads, encrypted or broken files are not charged. Set Maximum cost per run and the Actor stops before going over it, keeping everything already finished.

Error codes

Every failed document gets an item with status: "error" and one of these codes. One bad file never stops the rest of the batch.

Code	Meaning
DOWNLOAD_FAILED	The link returned an error (the HTTP status is included) or timed out
URL_NOT_ALLOWED	Not http(s), or points to a private network address
UNSUPPORTED_FORMAT	Not a PDF, DOCX or image (for example an HTML page, a ZIP, legacy .doc, or an Excel file)
FILE_TOO_LARGE	Over 100 MB
PDF_ENCRYPTED	Password-protected PDF
PDF_CORRUPT, DOCX_CORRUPT, IMAGE_CORRUPT	The file is damaged
PAGE_RANGE_INVALID	pages could not be parsed
OCR_LANGUAGE_UNAVAILABLE	Unknown OCR language code
NO_TEXT_FOUND	No text on any page (and OCR off, or the scan is unreadable)
MAX_COST_REACHED	Your maximum cost per run was reached

The run summary (OUTPUT in the key-value store) lists totals, the codes seen and the exact pages charged.

Good to know

• Tables: PDF tables are found from their drawn cell borders. Tables made only of aligned text (no lines) come through as text paragraphs.
• Columns: two-column pages are read left column first, then right. Complex magazine layouts may still mix.
• OCR quality depends on the scan. 200 dpi Tesseract works well on typed documents, poorly on handwriting.
• Not supported: password-protected PDFs, legacy .doc, PowerPoint, spreadsheets (use a spreadsheet tool), web pages.
• Typical speed: about 4 digital pages per second; about 1 OCR page per second.

Privacy

Files are processed in memory inside your own Apify run. Nothing is sent to any third-party service or AI model, and nothing is kept by the developer. Output lives in your run's storage under your account and retention settings.

Feedback

Found a document that converts badly? Open an issue on the Actor's Issues tab with a public link to it.

Part of the Errandkit catalog of data tools, APIs and MCP servers for AI agents: https://errandkit.com