PDF & Office to Markdown for RAG — tables, chunks, OCR

Turn PDFs, scanned documents, images, Word, Excel and PowerPoint files into clean Markdown that LLMs and RAG pipelines can use directly:

• 📊 Tables come out as tables — as Markdown, CSV and a cell array (rows), not as jumbled lines of text. Tables whose ruling only separates groups of rows are split back into one row per record.
• 📰 Correct reading order in two- and three-column layouts (papers, reports, government gazettes).
• 🧹 Repeated headers, footers and page numbers removed, and paragraphs broken across columns or pages joined back together.
• 🧩 Ready-to-embed chunks that follow section boundaries, carry their page range and section path (Introduction > Method), and never cut a table without repeating its header row.
• 🔍 Automatic OCR only where needed: pages without a usable text layer and images are OCR'd (Spanish + English by default; Portuguese, French, German and Italian available). Pages with text are not, so they stay fast and cheap.
• 📁 One Actor for every format: PDF, PNG, JPG, WEBP, TIFF, BMP, DOCX, XLSX, PPTX (including speaker notes), TXT, MD and CSV.

No LLM inside: deterministic, fast (≈0.2 s per text page) and your documents are not sent to any third-party AI service.

How to use it

1. Paste public URLs of your documents, or upload files and select the storage they were uploaded to in Storage with uploaded files (the Actor runs with limited permissions and can only read the storage you choose).
2. Pick the output: one item per document, per chunk (ready for a vector database) or per page.
3. Run it. Leave the input empty to try a demo document (not charged).

From code or an AI agent, call it through the Apify API or MCP with {"documents": ["https://example.com/report.pdf"]}.

Input

Field	Default	What it does
documents	—	Files or URLs (up to 50 MB each).
storage	—	The key-value store with your uploaded files.
output_mode	document	document, chunks or pages.
chunk_tokens	800	Target chunk size (≈4 characters per token). 0 = no chunks.
chunk_overlap	80	Text repeated at the start of the next chunk when a section is split by size.
ocr	auto	auto, always (PDFs with a broken text layer) or never.
ocr_languages	spa+eng	Languages of scanned pages.
pages	all	Page or slide range, e.g. 1-5, 8, 12-.
max_pages	200	Safety limit per document.
include_pages	false	Adds per-page Markdown, method and OCR confidence.
save_markdown_files	true	Also saves a .md file per document, with a link in the result.

Output (document mode)

{
  "file": "report.pdf",
  "status": "ok",
  "format": "pdf",
  "page_count": 12,
  "metadata": {"title": "Annual report", "author": "…", "creationdate": "2026-03-02T10:15:00+00:00"},
  "markdown": "# Annual report\n\n## Results\n\n| Plot | Water (m³) | Yield (kg) |\n| --- | --- | --- |\n| A1 | 1,200 | 3,450 |…",
  "tables": [{"table_id": 1, "page": 3, "n_rows": 4, "n_cols": 3, "rows": [["Plot", "Water (m³)", "Yield (kg)"], ["A1", "1,200", "3,450"]], "markdown": "…", "csv": "…"}],
  "chunks": [{"chunk_id": 1, "section": "Annual report > Results", "pages": [3, 4], "tokens_est": 612, "text": "…"}],
  "billing": {"page": 11, "ocr-page": 1},
  "stats": {"chars": 41250, "tokens_est": 10313, "n_tables": 4, "n_ocr_pages": 1, "removed_header_footer_lines": 22},
  "markdown_file": "https://api.apify.com/v2/key-value-stores/…/records/001-report.md",
  "warnings": []
}

In chunks mode each item is one chunk: file, chunk_id, chunk_count, section, pages, tokens_est, text, title. In pages mode each item is one page with its Markdown and tables.

Pricing

Pay per event, only for pages that produced text:

Event	Price	When
page	$0.001	A PDF page with a text layer, a slide, or every 3,000 characters of DOCX/XLSX/TXT output
ocr-page	$0.003	A scanned page or image that needed OCR and returned text

Empty pages, failed downloads, unsupported files and the demo are not charged. The run stops when it reaches the maximum cost per run you set.

Limits (honest list)

• Tables are detected on PDF pages with a text layer when they have ruling lines. Tables without lines and tables inside scanned pages come out as text (you get a warning).
• OCR is Tesseract: clean prints read very well; handwriting, very small text and photos with heavy blur do not.
• Headings in PDFs are inferred from font size and bold; Word headings come from their styles.
• Legacy .doc, .xls and .ppt are not supported (save them as DOCX/XLSX/PPTX or PDF). Password-protected PDFs are not supported.

Privacy and permissions

• Network access only to download the files you give it. The Apify token is only ever sent to api.apify.com.
• Files are processed in the run's container and deleted when it ends. Results stay in your own dataset and key-value store.
• Limited permissions: the Actor can only read the storage you select.
• Open-source engines with permissive licenses (pdfplumber, pypdfium2, Tesseract, OpenCV, python-docx, openpyxl, python-pptx). Pinned dependencies.

Feedback

Found a document that converts badly? Open an issue on the Actor page with a public link to it (or a similar one) and it will be looked at.