Document to Markdown: PDF, Word, PowerPoint, Excel & OCR converter

Convert PDF, Word, PowerPoint, Excel, HTML, EPUB, OpenDocument, RTF and scanned images into clean Markdown, plain text and ready-to-embed RAG chunks, for $2 per 1,000 documents.

📄 What is Document to Markdown?

It downloads the files you list (or reads them from your Apify storage) and returns one JSON row per document, with headings, tables, page numbers and metadata kept, and built-in Tesseract OCR for scans. All formats share the same output fields. No API key, no third-party AI service, and your files never leave the run.

Markdown an LLM can use: headings from font sizes and bold titles (in PDFs too), two-column papers in the right order, real Markdown tables (including booktabs tables without vertical lines), and repeated headers, footers and page numbers removed. Invisible <!-- Page 3 --> markers (Slide / Sheet for PowerPoint and Excel) and per-chunk pages and headingPath let you cite sources.

Use cases:

• AI / RAG pipelines: manuals, policies and contracts as chunks with page references for Pinecone, Qdrant, Weaviate or pgvector.
• LLM ingestion: clean Markdown instead of messy PDF text for GPT, Claude, Gemini or Llama, with fewer tokens.
• Developers and search teams: text and metadata (title, author, dates, pages) for Elasticsearch, Algolia or Typesense.
• Automation agencies: text from invoices, reports and scanned letters in n8n, Make or Zapier, or as an MCP tool for agents.
• Data migration: old DOC/RTF/ODT/XLS archives into Markdown for Notion, Confluence or Obsidian.

Output fields

Field	Description	Example
url / finalUrl	Your URL / the URL after redirects	https://arxiv.org/pdf/2401.18059
source	url or keyValueStore	url
keyValueStoreKey	Record key (store input)	null
fileName	File name	2401.18059v1.pdf
format	Detected format	pdf
mimeType	MIME type	application/pdf
sizeBytes	File size	2547113
title	Document title	RAPTOR: ...
pages	Total pages, slides, sheets or frames	23
pagesConverted	Pages converted (null for Word/HTML)	23
wordCount / charCount	Text size	12132
tableCount	Markdown tables	13
ocrUsed / ocrPages	OCR used, and on which pages	false / []
ocrConfidence	Mean Tesseract confidence, 0-100	95.8
markdown / text	Converted content (per output format)	<!-- Page 1 -->\n# RAPTOR...
pageContents[]	Optional per-page page, type, ocr, markdown/text	-
chunks[]	Optional index, text, headingPath, pages, charCount	see below
metadata	Author, creator, producer, dates...	{"producer": "pdfTeX-1.40.25"}
truncated	Pages skipped by your limit	false
warnings	Non-fatal notes	[]
error / errorCode	Failure reason (null on success)	null
convertedAt	UTC timestamp	2026-10-05T03:20:13Z

Supported formats

• Documents: PDF, DOCX/DOCM, PPTX, XLSX/XLSM, XLS, CSV/TSV, HTML, EPUB, ODT, ODS, ODP, RTF, DOC (legacy Word, text only), TXT, Markdown, JSON, XML and Jupyter notebooks.
• Images (OCR): PNG, JPG, TIFF (multi-page), BMP, GIF and WEBP.
• Not supported: legacy .ppt, ZIP archives and encrypted Office files. Save them as PPTX/DOCX or unzip them first.

Use cases: Developer tools (AI and LLM integrations, transform & process data, process & convert media) · Collect & distribute content.

🚀 How to use Document to Markdown

1. Open Document to Markdown on Apify Store and click Try for free.
2. Paste document URLs, one per line. You can also upload files or point the actor to an Apify key-value store that holds your files.
3. Choose the output format (Markdown, text or both). If you want them, turn on RAG chunks, per-page content or more OCR languages.
4. Click Start. The prefilled example (an arXiv paper, a Word file and a scanned TIFF) finishes in about 10 seconds.
5. Download the results, or read them through the API, n8n, Make, Zapier, LangChain or LlamaIndex.

🔧 Input guide

Document URLs (urls). Direct file links; the type is detected from the content.

• Good: https://arxiv.org/pdf/2401.18059, a pre-signed S3/Azure/GCS link, a key-value store record URL.
• Bad: a login-protected link, a Google Drive share link, or a web page that only links to the PDF (it is converted as HTML, with a warning).
• Duplicates are skipped; non-http(s) entries are listed in RUN_SUMMARY.

Uploads and key-value store (uploadedFiles, keyValueStoreId, keyValueStoreKeys). Uploads go to your own key-value store. For files already in Apify storage, give the store ID or name; all records are converted unless you list keys.

Output (outputFormat, includeTables, includePageMarkers, includePageContents). markdown (default), text or both. Tables and page markers are on by default. Per-page content adds pageContents[].

RAG chunks (chunkSize, chunkOverlap). 0 = off. Otherwise chunks[] of about N characters (minimum 200), split on headings, paragraphs and tables, with 200 characters of overlap by default. 1,500-3,000 characters suits most embedding models.

OCR (ocrMode, ocrLanguages). auto OCRs only images and pages without a text layer. always OCRs every PDF page (for broken text layers; every page is then charged). off never OCRs, so scans fail as free no_text errors. Languages: English, German, French, Spanish, Italian; each extra one slows OCR.

Limits and performance.

• maxPagesPerDocument: 200 (up to 5,000); extra pages are skipped and the item is truncated. pageRange: e.g. 1-5, 8, 20-.
• maxFileSizeMb: 50 (up to 500); larger files are skipped for free. Above 100 MB, use 4-8 GB of memory.
• Up to 10,000 documents per run and 10,000 rows per spreadsheet sheet.
• 5 parallel downloads, 120 s timeout, two retries on 429/5xx and timeouts. Each 2 GB of memory converts one more document in parallel (up to 4).

Other options. removeHeadersFooters (on) drops repeated PDF headers and footers. pdfPassword unlocks encrypted PDFs and is stored as a secret.

Polite downloading. robots.txt (including crawl-delay) is respected on every host and redirect. Private and internal network addresses are refused.

Full example input:

{
  "urls": [
    "https://arxiv.org/pdf/2401.18059",
    "https://raw.githubusercontent.com/microsoft/markitdown/main/packages/markitdown/tests/test_files/test.docx",
    "https://raw.githubusercontent.com/tesseract-ocr/test/main/testing/phototest.tif"
  ],
  "outputFormat": "markdown",
  "chunkSize": 2000,
  "chunkOverlap": 200,
  "ocrMode": "auto",
  "ocrLanguages": ["eng", "deu"],
  "maxPagesPerDocument": 200,
  "maxFileSizeMb": 50
}

📦 Output

One dataset item per document. Real item from a run on 2026-10-05 (markdown shortened):

{
  "url": "https://arxiv.org/pdf/2401.18059",
  "source": "url",
  "keyValueStoreKey": null,
  "fileName": "2401.18059v1.pdf",
  "format": "pdf",
  "mimeType": "application/pdf",
  "sizeBytes": 2547113,
  "title": "RAPTOR: RECURSIVE ABSTRACTIVE PROCESSING FOR TREE-ORGANIZED RETRIEVAL",
  "pages": 23,
  "pagesConverted": 23,
  "wordCount": 12132,
  "charCount": 75919,
  "tableCount": 13,
  "ocrUsed": false,
  "ocrPages": [],
  "ocrConfidence": null,
  "markdown": "<!-- Page 1 -->\n# RAPTOR: RECURSIVE ABSTRACTIVE PROCESSING FOR TREE-ORGANIZED RETRIEVAL\n\nParth Sarthi, Salman Abdullah, ...\n\n## ABSTRACT\n\nRetrieval-augmented language models can better adapt to changes in world state ...",
  "text": null,
  "metadata": { "creator": "LaTeX with hyperref", "producer": "pdfTeX-1.40.25", "createdAt": "2024-02-01T01:45:46", "modifiedAt": "2024-02-01T01:45:46" },
  "truncated": false,
  "warnings": [],
  "error": null,
  "errorCode": null,
  "convertedAt": "2026-10-05T03:20:13Z"
}

A table from the same paper, as it appears in markdown:

| Model | ROUGE | BLEU-1 | BLEU-4 | METEOR |
|---|---|---|---|---|
| SBERT with RAPTOR | 30.87% | 23.50% | 6.42% | 19.20% |
| SBERT without RAPTOR | 29.26% | 22.56% | 5.95% | 18.15% |

The scanned TIFF from the same run has "format": "image", "ocrUsed": true, "ocrConfidence": 95.8 and this markdown:

"This is a lot of 12 point text to test the ocr code and see if it works on all types of file format.\n\nThe quick brown dog jumped over the lazy fox. ..."

Optional chunks[] items (chunkSize: 2000, from the NIST AI RMF PDF, text shortened):

{"index": 10, "text": "uncertainties in their models and systems, which in turn can improve overall sys...", "headingPath": ["Artificial Intelligence Risk Management Framework (AI RMF 1.0)", "Part 1: Foundational Information", "1.1 Understanding and Addressing Risks, Impacts, and Harms"], "pages": [9, 10], "charCount": 1580}

Failed documents are free rows with error and errorCode: http, robots, too_large, password_required, unsupported_format, no_text, invalid_file, blocked_host, dns, timeout, empty_file (rarer: refused, tls, network, not_found, internal).

Dataset views: Documents (overview), Markdown, Metadata (MIME type, size, OCR details, warnings) and Errors.

RUN_SUMMARY (key-value store) holds the totals: documents, errors by code, formats, pages, OCR pages, charged events, invalid inputs, charge-limit status and duration. Your files are not saved.

Export as JSON, CSV, Excel, XML, RSS or an HTML table, or read it through the API.

💵 Pricing: $2 per 1,000 documents

Event	Price	When
document	$0.002	Per successfully converted document. Each started 20 pages (PDF pages, slides, sheets) counts once, so a 1-20 page file = 1, a 45-page file = 3. Word/HTML/text files count 500 words as one page (one event per 10,000 words)
ocr-page	$0.002	Per page or image actually read with OCR (auto mode never OCRs pages that have a text layer)

Apify also charges its tiny $0.00005 per-run start fee. Platform compute is included in the prices above. Failed, empty, password-protected and unsupported files are listed with the reason but never charged.

Typical run	Events	Cost
100 Word files or short PDFs (up to 20 pages)	100 document	$0.20
1,000 ten-page PDFs	1,000 document	$2.00
100 PDFs of 100 pages	500 document	$1.00
10 scanned PDFs of 20 pages	10 document + 200 ocr-page	$0.42

Add $0.00005 per run for the start fee. You can cap spending with Max charge per run. The run stops cleanly at your limit and never delivers a document it could not charge.

🔌 Use it via API

JavaScript (npm install apify-client):

import { ApifyClient } from 'apify-client';

const client = new ApifyClient({ token: 'YOUR_API_TOKEN' });
const run = await client.actor('kadi_bence/document-to-markdown').call({
    urls: ['https://arxiv.org/pdf/2401.18059'],
    chunkSize: 2000,
});
const { items } = await client.dataset(run.defaultDatasetId).listItems();
console.log(items[0].chunks.length);

Python (pip install apify-client):

from apify_client import ApifyClient

client = ApifyClient('YOUR_API_TOKEN')
run = client.actor('kadi_bence/document-to-markdown').call(
    run_input={'urls': ['https://arxiv.org/pdf/2401.18059'], 'outputFormat': 'both'}
)
for item in client.dataset(run['defaultDatasetId']).iterate_items():
    print(item['fileName'], item['wordCount'], item['errorCode'])

cURL (runs the actor and returns the items in one call):

curl -X POST "https://api.apify.com/v2/acts/kadi_bence~document-to-markdown/run-sync-get-dataset-items?token=YOUR_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"urls": ["https://arxiv.org/pdf/2401.18059"], "chunkSize": 2000}'

Apify CLI:

$apify call kadi_bence/document-to-markdown --input '{"urls": ["https://arxiv.org/pdf/2401.18059"]}'

MCP for AI agents (Claude, Cursor, VS Code and other MCP clients):

{"mcpServers": {"apify": {"url": "https://mcp.apify.com/?actors=kadi_bence/document-to-markdown"}}}

🔗 Integrations & scheduling

• Schedules: run on a cron, e.g. nightly.
• Webhooks: on "run succeeded", Apify calls your URL with the run ID.
• Zapier, Make, n8n: start runs and read results with the Apify app or node.
• Google Sheets, Slack, email: via Apify integrations or the tools above.

Recipes:

1. Convert new PDFs for a RAG pipeline. Your app or n8n/Make flow starts the actor with only the newly added PDF URLs and chunkSize: 2000 (it does not remember earlier runs). A "run succeeded" webhook calls your ingest endpoint, which embeds chunks[] and stores url, pages and headingPath as metadata.
2. Document inventory: export the Documents view to Google Sheets.
3. Failure alerts: a Make or n8n flow posts the Errors view to Slack.

🧰 Related tools by the same developer

Content pipelines

• Bulk Image Downloader: all images and files from pages into a ZIP. $0.80 / 1,000 files.

Website due diligence

• Website Tech Stack Detector: CMS, ecommerce, analytics and frameworks from 7,600+ fingerprints. $1.50 / 1,000 websites.
• SEO Site Audit Crawler: broken links, redirects, titles/meta, canonical and hreflang on every page. $1.50 / 1,000 pages.
• Website Screenshot & PDF API: full-page/mobile PNG, JPEG, WebP and PDF captures. $2.50 / 1,000 screenshots.
• Bulk WHOIS & RDAP Domain Lookup: registrar, age, expiry, DNS/MX/SPF/DMARC. $1.50 / 1,000 domains.

❓ FAQ

How does it work? It downloads each file, detects the format from the content and converts it with open-source libraries inside the run (pdfplumber for PDFs, Tesseract for scans). No external AI or OCR API is called.

What are the limits? 200 pages and 50 MB per document by default (up to 5,000 pages and 500 MB), 10,000 documents per run and 10,000 rows per sheet.

How good are PDF tables? Ruled tables and booktabs-style tables (scientific papers, many reports) become Markdown tables. Tables with no lines at all come out as text lines in reading order. Word, PowerPoint, Excel and HTML tables are always converted exactly.

Does OCR work on handwriting or photos? OCR is Tesseract 5.5 (LSTM). It works best on printed text and scans: typically 90%+ confidence on clean 200-300 dpi scans. Handwriting and poor phone photos give poor results.

How fast is it? In our tests: text PDFs about 0.1 s per page, OCR about 1-2 s per scanned page (1 CPU core), Office files under a second each. The prefilled 3-document example takes about 10 seconds.

Which URLs can it download? What about blocked sites? Direct file links you supply. Files disallowed by robots.txt fail as free robots errors. Login-protected files can't be fetched: use pre-signed links or upload them. Google Drive disallows automated downloads, so download and upload those files.

Are my files stored? Is it legal? Files stay in memory only for the conversion. Only your dataset and RUN_SUMMARY are saved, under your account's retention settings. Metadata such as author may contain personal names; handle it according to your own privacy obligations. You are responsible for having the right to process your documents (this is not legal advice).

Can I use it from code or an AI agent? Yes: API, Python/JS clients, CLI or MCP (see above), with the same fields for every format.

What does a big job cost? 10,000 ten-page PDFs cost $20; 1,000 PDFs of 100 pages cost $10, plus any OCR pages. Cap it with Max charge per run, maxPagesPerDocument or pageRange.

What if a document fails? It is listed with error and errorCode, is not charged, and the run continues. If a file should work, open an issue in the Issues tab with the URL and code.

What's new? See the changelog below.

📝 Changelog

• 1.0 (2026-10-05): First release. 20+ formats, PDF headings/tables/reading order, header and footer removal, page markers, RAG chunks with page references, Tesseract OCR (eng/deu/fra/spa/ita), key-value store input and file upload, password-protected PDFs. Failed documents are not charged.
• 2026-10-05: Store page rewritten.

🤝 Need a custom version?

I build custom scrapers, scheduled data feeds and integrations (CSV, Excel, Google Sheets, API, MCP for AI agents). Tell me the sites and fields you need:

• Email: bence.kadi@gmail.com
• Apify profile: https://apify.com/kadi_bence

Credits and licences

Built on open-source libraries, all under permissive licences:

• MarkItDown (MIT, © Microsoft) for Word, PowerPoint, HTML and EPUB
• pdfplumber (MIT) and pdfminer.six (MIT) for PDF text and layout
• pypdfium2 (Apache-2.0 / BSD-3) for page rendering
• Tesseract OCR (Apache-2.0) via pytesseract (Apache-2.0)
• openpyxl (MIT) and xlrd (BSD) for spreadsheets
• striprtf (BSD-3) for RTF
• antiword (GPL-2.0, run as a separate program) for legacy .doc

This actor is not affiliated with Microsoft, Adobe or Google. Product names are trademarks of their owners.

Sample documents in the prefilled input:

• the arXiv paper 2401.18059 (CC BY 4.0, Sarthi et al.)
• test.docx from the MarkItDown test suite (MIT)
• phototest.tif from the Tesseract test repository (Apache-2.0)

Related Actors

More low-cost Actors by the same developer, built on official APIs and public data:

• Website Screenshot & PDF API — full-page screenshots and PDFs of any URL
• Bulk Image Downloader — download all images from URLs as a ZIP
• AI Visibility Tracker — how ChatGPT, Perplexity, Gemini and Claude mention your brand
• TED Tenders Scraper — EU public tenders and contract award winners
• SAM.gov Scraper — US federal contract opportunities and awards
• SEO Site Audit — 0-100 SEO score per page, broken links and fix hints
• Website Tech Stack Detector — CMS, ecommerce, analytics and frameworks of any website

All my Actors: apify.com/kadi_bence