Documents to Markdown for RAG (Docling)

Turn PDF and DOCX files into clean, structured Markdown ready for LLMs, RAG pipelines and vector databases. Conversion runs on Docling, the open-source document parser from IBM Research (now under the LF AI & Data Foundation), which understands page layout, reading order, headings, lists and tables instead of dumping raw text.

What you get per document

• markdown: the full document as Markdown, headings and tables included
• pageCount, title, fileName, format (pdf or docx), fileSizeBytes
• tableCount and tables: each table with its page, number of rows and columns, and the table as Markdown
• chunks (optional): the Markdown split into heading-aware pieces for embeddings, each with its heading path
• billedPages, markdownChars, processingSecs, status (ok or partial) and warnings

Input

Field	Description
documentUrls	Public links to PDF or DOCX files, up to 100 per run. The format is detected from the file bytes, so links without an extension work.
uploadedFile	Optional file uploaded from your computer, converted together with the URLs.
chunking	Also return RAG chunks (default false).
chunkSize / chunkOverlap	Maximum characters per chunk (default 2,000) and characters repeated from the previous chunk (default 200).
tableStructure	Rebuild PDF tables cell by cell with Docling's TableFormer model (default true).
includeTables	Include each table's Markdown in tables (default true).
ocr	OCR scanned PDFs (default false; digital PDFs already contain their text).
maxPagesPerDocument / maxFileSizeMb	Safety limits (default 200 pages, 50 MB). Documents over the limits return an error and are free.

The default input converts one public page from a research paper with a table, so you can see the output in under a minute.

Output example

{
  "url": "https://example.com/annual-report.pdf",
  "fileName": "annual-report.pdf",
  "format": "pdf",
  "status": "ok",
  "title": "Annual Report 2025",
  "pageCount": 12,
  "tableCount": 3,
  "tables": [{"page": 4, "rows": 8, "columns": 5, "markdown": "| Year | Revenue | ... |"}],
  "markdown": "# Annual Report 2025\n\n## Highlights\n\n...",
  "chunks": [{"index": 0, "headings": ["Annual Report 2025", "Highlights"], "text": "...", "chars": 1840}],
  "billedPages": 12,
  "markdownChars": 38211,
  "processingSecs": 41.3
}

Files that cannot be downloaded, are not PDF or DOCX, or fail to convert return an item with status: "error" and an error message, and you are not charged for them.

Chunking for RAG

Chunks follow the document structure: sections are cut at headings, small sibling sections are merged up to chunkSize, and long sections are split at paragraph, then line, then word boundaries. Each chunk carries its headings path (for example ["User guide", "Installation"]), which you can store as metadata or prepend to the text before embedding.

How much does it cost?

Pay per event, no usage fees on top: US$ 0.004 per converted page, plus a tiny start charge of US$ 0.00005 per run.

DOCX files have no fixed pages, so every 3,000 characters of Markdown count as one page (minimum 1). Documents that fail are free. A 1-page PDF costs about US$ 0.004, a 5-page one US$ 0.02 and a 100-page one US$ 0.40. Put several documents in one run to pay the start only once.

Speed on Apify (measured on 30/10/2026): the default 1-page input took about 18 s, of which roughly 15 s is fixed start-up (loading the Docling models); each extra PDF page adds roughly 0.2 to 0.5 s. A 100-page PDF converted in about 24 s on one thread.

The niche leader charges about US$ 0.02 per document regardless of size; we are cheaper for documents up to 5 pages and charge in proportion for longer ones.

$0.004 per page (page event, pay per event). Charged per page of each converted document (DOCX: each 3,000 Markdown characters count as one page). Failed documents are free.

• apify-actor-start: $0.00005 (Charged once when a run starts (Apify synthetic event).)

You get	Free tier	Bronze tier (10% off)	Gold tier (30% off)
100 pages	$0.40	$0.36	$0.28
1,000 pages	$4.00	$3.60	$2.80
10,000 pages	$40.00	$36.00	$28.00

Discounts follow your Apify subscription tier: Bronze 10%, Silver 20%, Gold and above 30%. No subscription to this Actor and no charge for platform usage: you pay only for results. The Apify Free plan includes $5 of monthly credit, enough for about 1,250 pages at the list price. You can cap the cost of any run with Maximum cost per run; the Actor stops when it is reached.

Data source and limits

Your own documents, or public links you provide. Files are processed in the run's memory and are not stored anywhere else; the dataset of the run keeps the converted Markdown. Conversion uses Docling's open models (layout analysis and TableFormer), which run inside the Actor, with no external AI API. Scanned PDFs need ocr: true. Complex layouts (multi-column magazines, forms) are converted on a best-effort basis.

Docling is MIT licensed; its models are released under permissive licenses. This Actor is not affiliated with IBM or the Docling project.

More questions

Which file formats are supported?

PDF and DOCX. The format is detected from the first bytes of the file, so links without an extension or with a wrong content type still work. Other files (PPTX, XLSX, HTML pages, images, ZIP) return an item with status: "error" and Not a PDF or DOCX file, and are not charged.

Does it work with scanned PDFs?

Yes, with ocr: true. OCR is off by default because digital PDFs already contain their text and OCR makes the run slower. If a scanned PDF comes back with almost no Markdown, run it again with OCR on.

What chunk size should I use for embeddings?

chunkSize is in characters, overlap included; about 4 characters make one token in English, so the default of 2,000 is roughly 500 tokens. Chunks are cut at headings first, small sibling sections are merged, and long sections are split at paragraph, line and then word boundaries. chunkOverlap (default 200) repeats the end of the previous chunk of the same section and is capped at half the chunk size. Each chunk has a headings path you can keep as metadata or prepend before embedding.

Are tables kept as tables?

Yes. Tables are rebuilt cell by cell with Docling's TableFormer model and written as Markdown tables, both inside markdown and in tables with their page, rows and columns. Turn off tableStructure for faster runs on text-only documents, or includeTables to keep only each table's page and size in tables.

Can I send a Google Drive or Dropbox share link?

Only if the link downloads the file itself. Share pages return HTML, which is rejected as not a PDF or DOCX (free). Use a direct download link, or upload the file in uploadedFile.

What are the limits per run?

Up to 100 documents per run (repeated URLs are converted once), processed one at a time. maxPagesPerDocument (default 200, up to 2,000) and maxFileSizeMb (default 50, up to 500) skip larger documents with an error, free of charge. When a run reaches your Maximum cost per run, it stops after the current document.

Are my documents sent to an AI service or stored?

No external AI API is called: Docling's layout and table models run inside the Actor. Files are downloaded into the run's memory and only the converted result is saved, in the run's dataset, which follows your Apify storage retention.

Is a partly broken PDF charged?

When Docling converts only part of a document, the item comes back with status: "partial" and the problems in warnings, and its pages are charged like a normal result. Documents that fail completely are free.

Use it via API

Run it from any language and get the results as JSON. Replace <YOUR_API_TOKEN> with your Apify API token.

curl -X POST "https://api.apify.com/v2/acts/br_data_hub~docs-to-markdown/run-sync-get-dataset-items?token=<YOUR_API_TOKEN>" \
  -H "Content-Type: application/json" \
  -d '{"documentUrls": ["https://raw.githubusercontent.com/docling-project/docling/v2.145.0/tests/data/pdf/sources/2305.03393v1-pg9.pdf"]}'

from apify_client import ApifyClient

client = ApifyClient("<YOUR_API_TOKEN>")
run = client.actor("br_data_hub/docs-to-markdown").call(run_input={'documentUrls': ['https://raw.githubusercontent.com/docling-project/docling/v2.145.0/tests/data/pdf/sources/2305.03393v1-pg9.pdf']})
for item in client.dataset(run["defaultDatasetId"]).iterate_items():
    print(item)

You can also schedule runs, send results to webhooks, or connect the Actor to Make, Zapier, n8n, Google Sheets and other integrations.

Use it with AI agents (MCP)

This Actor is ready for AI agents: typed input and output schemas, pay-per-event pricing, limited permissions and short runs.

• Apify MCP server: add https://mcp.apify.com/?tools=br_data_hub/docs-to-markdown to Claude, ChatGPT, Cursor, VS Code or any MCP client, and the agent can call this Actor as a tool.
• Agent frameworks: call it from LangChain, LlamaIndex, CrewAI or the OpenAI Agents SDK through the Apify integrations, or with the API call above.
• Results come as structured JSON with stable field names, documented in the Output tab.

FAQ

Where does the data come from?

Your own documents, or public links you provide. Files are processed in the run's memory and are not stored anywhere else; the dataset of the run keeps the converted Markdown. Conversion uses Docling's open models (layout analysis and TableFormer), which run inside the Actor, with no external AI API. Scanned PDFs need ocr: true. Complex layouts (multi-column magazines, forms) are converted on a best-effort basis.

How much does it cost?

$0.004 per page. Discounts follow your Apify subscription tier: Bronze 10%, Silver 20%, Gold and above 30%. See How much does it cost? above for examples.

Do I pay for errors or empty results?

No. Charged per page of each converted document (DOCX: each 3,000 Markdown characters count as one page). Failed documents are free.

Can I run it on a schedule or connect it to other tools?

Yes. Use Apify schedules to run it daily or weekly, webhooks to get notified, and integrations such as Make, Zapier, n8n and Google Sheets. Everything also works through the API.

Can AI agents use it?

Yes. It works as a tool in the Apify MCP server and in agent frameworks, and AI agents can pay per event. See Use it with AI agents (MCP).

Is it legal to use this data?

The Actor reads only public sources. You are responsible for how you use the data and for complying with the source's terms and with privacy laws such as GDPR.

I found a bug or need a feature. What should I do?

Open an issue in the Issues tab. We answer within 14 days, usually much sooner.

More Actors from BR Data Hub

Actor	What it does
Website to llms.txt Generator	Crawl one website (robots.txt and sitemap aware), generate spec-compliant llms.txt and llms-full.txt files, au
RSS & Atom Feed Reader with Full Text	Read RSS, Atom and JSON Feed (or find the feed of any website) and get clean items with ISO dates, full articl
Research Papers Search for RAG, from OpenAlex data	Search 250M+ scholarly papers from OpenAlex data, or look up DOIs: authors with institutions, venue, citations
Brazil CNPJ Lookup API (Consulta CNPJ)	Look up Brazilian companies by CNPJ: registration status, partners (masked CPF), activities, tax regime and sc

See all our Actors at apify.com/br_data_hub.

Usar em português

Informe links de arquivos PDF ou DOCX (ou envie um arquivo) e receba o documento em Markdown limpo, com número de páginas, tabelas detectadas e metadados. Ative chunking para receber o texto já dividido em pedaços por seção, prontos para RAG. A cobrança é por página convertida; arquivos com erro não são cobrados.

Quanto custa: $0.004 por resultado (page); 1.000 resultados saem por $4.00. Desconto pelo nível da assinatura Apify: Bronze 10%, Silver 20%, Gold ou acima 30%. Erros e consultas sem resultado não são cobrados. Use também pela API ou por agentes de IA no servidor MCP da Apify (https://mcp.apify.com/?tools=br_data_hub/docs-to-markdown).