PDF to Markdown Extractor & Document Parser: Tables, JSON, OCR

What does PDF to Markdown Extractor & Document Parser do?

Document to Markdown, Tables & JSON converts PDF, Word (DOCX), Excel (XLSX), PowerPoint (PPTX), HTML pages, legacy Excel (XLS), CSV, TXT and images into clean, LLM-ready Markdown, structured tables (JSON + CSV), plain text, metadata and RAG chunks. Scanned PDFs and images are read with OCR in 7 languages.

Use it as a PDF extractor, PDF table extractor (PDF to Excel/CSV/JSON), document parser for RAG, DOCX/PPTX to Markdown converter or OCR API, all in one Actor.

Give it a list of document URLs and get one structured JSON record per document. No API key and no setup. You pay only for documents that are processed successfully.

It runs on the Apify platform, so you also get an API, scheduling, integrations (Make, Zapier, n8n, LangChain, LlamaIndex), webhooks, monitoring, and access for AI agents through the Apify MCP server.

Why use it?

• 🤖 RAG and AI agents: turn documents into Markdown with preserved headings, lists and tables, pre-split into chunks with page numbers.
• 📊 Table extraction: every table in a PDF, Word, PowerPoint or Excel file comes back as header + rows + ready-to-use csv, in its original position in the Markdown. This includes academic-paper tables drawn with horizontal rules only (LaTeX booktabs), which most PDF tools return as jumbled text. Tables that continue over several pages (with the header repeated on each page) are merged into one table.
• 🧾 Scanned documents: OCR runs only on pages that need it (auto mode), so native PDFs stay fast and cheap.
• 📰 Clean reading order: two-column layouts (papers, journals, forms) are read column by column, and running headers, footers and page numbers are removed, so RAG chunks are not polluted.
• 🧠 AI field extraction (optional): ask for specific fields (invoice number, total, dates, parties, line items…) and get them as clean JSON matching your schema, powered by Claude with your own Anthropic API key.
• 🔗 Paste any link: Google Drive, Google Docs/Sheets/Slides, Dropbox, OneDrive/SharePoint and GitHub share links are converted to downloads automatically.
• 🕸️ Whole document libraries: point it at a web page (reports, filings, publications) and enable followDocumentLinks to convert every PDF/Office file linked from it.
• ✍️ Faithful Markdown: bullet and numbered lists and bold sub-headings in PDFs; bold, italic, links and nested lists in Word; chart data from PowerPoint as tables; OCR text re-flowed into real paragraphs.
• 🗂️ One tool, many formats: the file type is detected from the content, not the URL, so links without extensions work.
• 🛡️ Predictable cost: failed downloads and unsupported files are never billed, and the run stops cleanly when your spending limit is reached.

How to convert a PDF to Markdown and tables

1. Click Try for free.
2. Paste one or more document URLs into Document URLs.
3. Choose what to return: Markdown, tables, text, per-page content and/or RAG chunks.
4. Click Start and download the results as JSON, CSV, Excel or HTML, or fetch them via API.

Input

Field	Description	Default
sources	Document URLs (PDF, DOCX, XLSX, XLS, PPTX, HTML, CSV, TXT, PNG, JPG, TIFF, BMP, GIF, WEBP)	—
base64Files	Files without a URL: [{"fileName": "a.pdf", "content": "<base64>"}] (ideal for AI agents)	—
outputs	Any of markdown, tables, text, pages, chunks	markdown, tables
followDocumentLinks / maxLinkedDocuments	Convert the documents linked from a web page instead of the page itself	false / 20
saveFiles	Also save each document as a downloadable .md file, each table as a .csv file and all tables as an Excel workbook (.xlsx, one sheet per table); links in files	false
pdfPassword	Password for encrypted PDFs (stored as a secret)	—
maxCharactersPerDocument	Truncate text for AI agents with small context windows (0 = no limit)	0
removeHeadersFooters	Drop lines repeated at the top/bottom of most pages	true
tableDetection	lines (ruled tables, precise) or text (also borderless tables, experimental)	lines
ocr	auto (only scanned pages), always, never	auto
ocrLanguages	eng, spa, deu, fra, por, ita, nld	eng
pageRange	e.g. 1-5, 8 (PDF pages, slides or Excel sheets)	all
chunkSize / chunkOverlap	RAG chunk size and overlap in characters	1000 / 100
maxPages, maxFileSizeMb, maxConcurrency	Safety limits	500, 50, 2

Example input:

{
  "sources": [{ "url": "https://example.com/annual-report.pdf" }],
  "outputs": ["markdown", "tables", "chunks"],
  "ocr": "auto",
  "ocrLanguages": ["eng", "spa"]
}

Output

One record per document. You can download the dataset in various formats such as JSON, HTML, CSV, or Excel.

{
  "source": "https://example.com/annual-report.pdf",
  "status": "ok",
  "fileType": "pdf",
  "metadata": { "title": "Annual Report", "author": "ACME", "created": "2026-03-01T10:00:00", "pageCount": 12 },
  "stats": { "pagesProcessed": 12, "ocrPages": 0, "tables": 3, "characters": 18234 },
  "markdown": "# Annual Report\n\nThis report summarizes...\n\n| Year | Revenue | Region |\n| --- | --- | --- |\n| 2024 | 120 | EU |",
  "tables": [
    { "index": 0, "page": 1, "header": ["Year", "Revenue", "Region"], "rows": [["2024", "120", "EU"]], "csv": "Year,Revenue,Region\n2024,120,EU\n" }
  ],
  "chunks": [{ "index": 0, "text": "# Annual Report\n\nThis report...", "pageStart": 1, "pageEnd": 2, "tokens": 245 }],
  "warnings": []
}

Failed documents return "status": "error" with an error message and are not billed.

Data fields

Field	Description
markdown	Whole document as Markdown: headings, paragraphs, lists, tables in reading order
tables	Each table as header, rows, csv, with its page/slide/sheet number (pageEnd for tables merged across pages)
text	Plain text
pages	Markdown per PDF page, slide or sheet, with an ocr flag
chunks	RAG chunks that respect paragraph boundaries, with pageStart/pageEnd and an approximate tokens count
metadata	Title, author, subject, creation/modification dates, page count, sheet names, detected language (ISO code)
stats	Pages processed, OCR pages, tables found, characters
warnings	Anything worth knowing (skipped pages, scanned pages with OCR disabled…)

How much does it cost to convert documents?

Pay-per-event pricing, with no monthly fee:

Event	Price
Run start (per GB of memory, default 2 GB)	$0.001
Document processed successfully	$0.003 (about $3 per 1,000 documents)
Page read with OCR (scanned pages and images only)	$0.01
AI extraction (only with extractionFields/extractionSchema; Claude usage is billed to your own key)	$0.005

A run converting 100 native PDFs costs $0.002 + 100 × $0.003 = $0.302. A 5-page scanned PDF costs $0.002 + $0.003 + 5 × $0.01 = $0.055. Put many documents in one run to spread the start fee. Set Max cost per run in the run options, and the Actor stops cleanly when the limit is reached.

Extract invoice, contract or form fields with AI

Add extractionFields (or a full extractionSchema) and your Anthropic API key. Each document then gets an extracted object with exactly those fields; missing values are null.

{
  "sources": [{ "url": "https://example.com/invoice-1043.pdf" }],
  "extractionFields": ["invoice_number", "invoice_date", "vendor_name", { "name": "total", "type": "number" }],
  "extractionInstructions": "Dates as YYYY-MM-DD",
  "anthropicApiKey": "sk-ant-...",
  "extractionModel": "claude-sonnet-5"
}

"extracted": { "invoice_number": "1043", "invoice_date": "2026-08-31", "vendor_name": "ACME Corp", "total": 1250.0 }

• Works on every supported format, including scanned PDFs (OCR runs first).
• Claude usage is billed by Anthropic to your key; this Actor adds one small AI extraction event per successful extraction.
• Use extractionSchema for nested data such as line items ("type": "array" of objects).

Real-time API (Standby mode) for AI agents and apps

Need an answer in about a second instead of waiting for a run to start? Call the Actor's Standby endpoint. It stays warm and responds over plain HTTP, with the same pricing:

$curl "https://<username>--document-to-markdown-tables.apify.actor/?url=https://example.com/report.pdf&outputs=markdown,tables"   -H "Authorization: Bearer <YOUR_APIFY_TOKEN>"

• GET /?url=...&outputs=...&ocr=...&pageRange=... returns one result object.
• POST / with the same JSON as the normal input (sources, base64Files, options) returns {"results": [...]}.

The exact URL is shown in the Actor's Standby tab.

Use it with AI agents (MCP): Claude, ChatGPT, Cursor

This Actor works as a tool for AI agents through the Apify MCP server. Add it to Claude Desktop, Claude Code, Cursor, VS Code or any MCP client with this server URL:

https://mcp.apify.com?tools=fguiraud/document-to-markdown-tables

Then just ask in plain language, for example:

• "Convert this PDF to Markdown and give me its tables: https://example.com/annual-report.pdf"
• "Read these three invoices and return the invoice number, date and total of each"
• "Find all PDF reports linked on this page and summarise them"

Smallest useful input for an agent:

{
  "sources": [{ "url": "https://example.com/report.pdf" }],
  "outputs": ["markdown", "tables"],
  "maxCharactersPerDocument": 50000
}

• maxCharactersPerDocument keeps long documents inside the model's context window.
• base64Files lets an agent send a file it already has, without a public URL.
• Need a reply in about a second? Use the Standby endpoint above.

Agents pay per result like everyone else (Apify also supports agentic payments), so there is no subscription to set up.

Tips

• Leave OCR on auto: it runs only on pages without a text layer.
• Use pageRange to process only the pages you need from large reports.
• Large batches are safe: downloads that fail with a temporary error (5xx, 429, network) are retried, and if the platform restarts or migrates the run, documents already processed are skipped, so they are never charged twice.
• For RAG, select chunks and tune chunkSize to your embedding model (e.g. 800–1500 characters).
• Tables with grid lines and tables with only horizontal rules (typical in scientific papers) are detected automatically. For tables with no lines at all, try tableDetection: text: it finds more tables, but may split columns imperfectly.

FAQ and limitations

• Which files are not supported? Legacy Word and PowerPoint files (.doc, .ppt): save them as DOCX or PPTX first. Legacy Excel .xls and password-protected PDFs (with pdfPassword) are supported.
• Private Google Drive files? The file must be shared as "Anyone with the link"; otherwise it is reported as not accessible and is not billed.
• Web pages: static HTML is converted to Markdown with navigation, footers and cookie banners removed. JavaScript is not executed, so pages that render their content in the browser may come back mostly empty.
• Layouts: one- and two-column pages are supported. Three or more columns, and text inside figures, may come out in an imperfect order. Rotated text (for example margin stamps) is skipped.
• Can I process local files? Yes: send them in base64Files (up to ~9 MB of input per run), or upload them anywhere reachable by URL and use sources.
• Privacy: documents are processed in memory during the run and are not stored anywhere except in your own run's dataset.

Found a bug or need a new format? Open an issue in the Issues tab. Feedback is welcome.