Document to Markdown — PDF, Word, PowerPoint, Excel, OCR

Convert documents to clean Markdown for LLMs, RAG pipelines and AI agents: PDF, Word (DOCX), PowerPoint (PPTX), Excel (XLSX), HTML, CSV, JSON, XML, EPUB, text and images. Give it file URLs, get one Markdown result per file: headings, paragraphs, lists and tables as real Markdown tables. Scanned PDF pages and images are read with OCR automatically, and the output can come pre-split into RAG chunks with page numbers.

$2 per 1,000 pages ($1.80, $1.60 and $1.40 on the Starter, Scale and Business plans). $5 per 1,000 pages read with OCR ($4.50, $4.00, $3.50). Files that fail are free. No subscription, no platform usage fees. (Plus a tiny start fee of $0.00005 per GB of run memory: $0.0001 per run at the default 2 GB.)

Quick start

Press Start with the example input. It converts a sample PDF and a sample Word file:

{
  "urls": [
    "https://raw.githubusercontent.com/microsoft/markitdown/main/packages/markitdown/tests/test_files/test.pdf",
    "https://raw.githubusercontent.com/microsoft/markitdown/main/packages/markitdown/tests/test_files/test.docx"
  ]
}

Each file becomes one dataset row (trimmed here):

{
  "url": "https://.../test.pdf",
  "fileName": "test.pdf",
  "format": "pdf",
  "createdAt": "2023-10-05T00:07:56",
  "totalPages": 1,
  "pages": 1,
  "ocrPages": 0,
  "tables": 0,
  "wordCount": 774,
  "chars": 5191,
  "markdown": "<!-- page 1 -->\n\n1\n\n### Introduction\n\nLarge language models (LLMs) are becoming a crucial building block ..."
}

What it does

• PDF: the text layer page by page, with wrapped lines joined into paragraphs, hyphenation undone, bullet lists kept and headings detected from the font size (#, ##, ###). <!-- page N --> markers let you cite page numbers.
• PDF tables: tables with lines between cells and borderless tables (columns aligned by spacing, as in most reports and financial statements) become Markdown tables in their place in the text, with the caption above. Rows that a PDF packs into one cell are split back into rows, and the table's text is not repeated as loose numbers. Typically under a second per page, on CPU.
• RAG chunks (optional): set RAG chunk size and each result also gets chunks, each with its text, page range and nearest heading. Paragraphs and tables are not cut in the middle; big tables are split between rows with the header repeated.
• Metadata: title, author and creation date from PDFs and Office files, plus page, table and word counts.
• Scanned PDFs and images: pages without a text layer are read with Tesseract OCR and rebuilt as Markdown: paragraphs instead of broken lines, bullet lists as lists, large lines as headings. Choose the language (eng, deu, fra, spa, ita, por, nld, tur, pol, swe, or several with +).
• Word: headings, lists, tables and links as Markdown.
• PowerPoint: one section per slide with its title, text, tables and notes.
• Excel: one Markdown table per sheet.
• HTML, CSV, JSON, XML, EPUB, TXT, MD: converted or kept as readable Markdown.
• RAG friendly: turn on One result per page to get one item per PDF page, ready for chunking and embedding.

The file type is detected from the file itself, so links without an extension work too.

Example: a table from a PDF report

From this NREL report page, where plain text extraction gives a column of loose numbers:

Table 2-1. Simulated fuel savings from isolated cycle improvements

| Cycle Name | KI (1/km) | Distance (mi) | Percent Fuel Savings |  |  |  |
|---|---|---|---|---|---|---|
|  |  |  | Improved Speed | Decreased Accel | Eliminate Stops | Decreased Idle |
| 2012_2 | 3.30 | 1.3 | 5.9% | 9.5% | 29.2% | 17.4% |
| 2145_1 | 0.68 | 11.2 | 2.4% | 0.1% | 9.5% | 2.7% |
| 4234_1 | 0.59 | 58.7 | 8.5% | 1.3% | 8.5% | 3.3% |

Input

Field	Meaning
urls	Links to the files. Public http(s) URLs, up to the size limit. The same link twice is converted once.
extractTables	Output PDF tables as Markdown tables (default on).
ocr	auto (default): OCR a PDF page only when it has no text but has an image. always: OCR every PDF page. never: text layer only.
ocrLanguage	Language of scanned text, for example eng or eng+deu.
maxPages	Stop after this many pages of each PDF or image (0 = all). Caps the cost of long files.
splitByPage	One result per PDF page instead of one per file.
pageMarkers	Insert <!-- page N --> before each PDF page (default on).
chunkSize, chunkOverlap	Also return RAG chunks of about this many characters (0 = off), with this much overlap (default 200).
requestHeaders	Extra HTTP headers, for example {"Authorization": "Bearer …"} for files behind a token. Sent only to the host of each URL, never to a redirect on another host.
maxFileSizeMb	Skip files larger than this (default 100 MB, up to 500).
maxConcurrency	Files converted in parallel (default 3, up to 10).

Output

Field	Meaning
url, finalUrl	The link you gave, and where it redirected to (if it did)
fileName, format	File name and detected type
title, author, createdAt	Document metadata, when the file has it
totalPages, pages	Pages in the PDF, and pages converted (and charged)
ocrPages	Pages read with OCR
tables	Tables found and output as Markdown tables
wordCount	Words in the Markdown
chunks	With RAG chunk size: index, text, chars, pageStart, pageEnd, heading
unit	For Office files: slide, sheet or page (see pricing)
markdown, chars	The Markdown and its length
markdownUrl, markdownTruncated	For very long documents (over 1.5 million characters) the full Markdown is saved as a file and linked here
page, ocr	With One result per page: the page number and whether it was read with OCR
stoppedAtSpendingLimit	true if the run's spending limit was reached in this file
error	Why a file failed (not found, too large, password-protected, unsupported type)

What counts as a page

File	One billable page is
PDF	one page (OCR pages are charged as OCR pages instead)
Image	one image, or one frame of a multi-page TIFF (always OCR)
PowerPoint	one slide
Excel	one sheet
Word, HTML, EPUB, CSV, JSON, XML, text	3,000 characters of output, at least one

What it costs

Run	Price (no discount)
100 PDFs of 10 pages with a text layer	$2.00
1,000 slides	$2.00
200 scanned pages (OCR)	$1.00
A 300-page report, capped with maxPages: 50	$0.10

Good to know

• No per-run file limits on any Apify plan, including the free plan: convert as many files as your budget allows.
• You only pay for pages you get. Failed downloads, unsupported files and password-protected PDFs are free.
• The run stops cleanly at your spending limit, and returns the pages that were paid for.
• Private and internal addresses (localhost, 10.x, 192.168.x, cloud metadata) are blocked, including every redirect.
• Files behind a login are not supported: the link must download the file directly.
• Borderless tables are found where numbers line up in columns; tables of mostly long text without lines come out as paragraphs. Merged header cells are left empty in the Markdown table. Tables inside scanned pages come out as text.
• Multi-column pages come out as readable text in reading order, not as a pixel-perfect copy.
• Failed downloads (connection errors, HTTP 429 and 5xx) are retried twice before the file is reported as failed.
• You are responsible for converting files you have the right to use.

FAQ

Can I use it in a RAG pipeline? Yes. Set RAG chunk size (for example 1500) and embed each item's chunks[].text, keeping url, pageStart and heading as metadata for citations. Chunking is free: you pay for pages, not chunks.

Are tables kept? Yes, PDF tables (with or without lines) and all tables in Word, PowerPoint, Excel and HTML come out as Markdown tables. Set extractTables to false if you prefer plain text.

How do I convert files behind a login or token? Use a pre-signed URL (S3, Google Cloud Storage, Azure) or pass the token in requestHeaders.

How do I convert files I have locally? Upload them anywhere that gives a direct download link (for example an Apify key-value store, S3 or a public bucket) and pass the links.

Can an AI agent use it? Yes, through the Apify MCP server: the agent passes file URLs and gets Markdown back.

Can I run it from Make, Zapier, n8n or my own code? Yes, like any Apify Actor: through Apify's integrations or the Apify API.

More from this developer

• Bulk HTTP Requests: call any API once for every row of a CSV, JSON list or Apify dataset.
• Website Screenshot + Change Detection: screenshots of many URLs on desktop and mobile, with visual change detection between runs.
• Keyword Search Volume: Google search volume, CPC and trends for a keyword list, in up to 20 countries in one run.
• Audio to Text: audio, video and podcast feeds to timestamped text, SRT/VTT subtitles and RAG chunks with Whisper, $5 per 1,000 audio minutes.
• Video to Scenes: scene cuts, one keyframe per scene, a contact sheet and the transcript of each scene, $2 per 1,000 frames.
• Career Site Jobs: open jobs from Greenhouse, Lever, Ashby, Workable, Recruitee and Workday career sites, with a new-jobs monitor and a hiring summary per company, $1 per 1,000 jobs.

Questions or a missing feature: open an issue on the Actor page.