Docling Document Converter: PDF, Word, PowerPoint & HTML to Markdown/JSON for LLMs

Turn PDFs (including scanned ones), DOCX, PPTX, XLSX, HTML, Markdown, CSV and images into clean, LLM-ready Markdown or the full structured Docling JSON. Headings, reading order, lists and tables are kept. You can also have documents split into RAG chunks with heading paths and page numbers.

It runs IBM's open-source Docling (MIT) with its layout, TableFormer and OCR models built into the image. Your documents never leave the Actor run: no third-party APIs and no LLM calls.

By Tallow Lane Labs

What you get

Markdown	Real # headings, ordered/bulleted lists, pipe tables, and <!-- image --> placeholders where figures sit.
Docling JSON	The lossless DoclingDocument: every text, table cell and picture, with page number and bounding box.
RAG chunks (optional)	Token-sized chunks (Docling HybridChunker) that respect sections, tables and lists. Each chunk has headings, pageNumbers and tokenCount. Chunks are delivered in the document row, or as one dataset row per chunk ready for vector DB integrations.
OCR	Scans and image-only pages are OCR'd automatically (RapidOCR / PP-OCRv6). You can force OCR or turn it off.
Tables	TableFormer fast (default) or opt-in accurate for dense tables. Tables in Office and HTML files are read natively.

Supported inputs

• Files: PDF, DOCX, PPTX, XLSX, HTML, MD, CSV, AsciiDoc, PNG, JPG, TIFF, BMP and WEBP. Up to 500 MB each (100 MB by default).
• Two ways in: upload files in the input form (Upload documents), or paste direct links (Document URLs). You can mix them in one run, with up to 1,000 documents per run.
• Not supported: password-protected PDFs, legacy .doc and .ppt files, and URLs that need a login. These come back as a failed row and are not charged.

Input example

{
  "urls": [{ "url": "https://arxiv.org/pdf/2408.09869" }],
  "outputFormat": "both",
  "ocrMode": "auto",
  "tableMode": "fast",
  "maxPagesPerDocument": 200,
  "chunking": true,
  "chunkMaxTokens": 512
}

Output

You get one dataset row per document (recordType: "document"):

{
  "fileName": "2206.01062.pdf", "format": "pdf", "status": "success",
  "title": "DocLayNet: A Large Human-Annotated Dataset for Document-Layout Analysis",
  "pagesConverted": 9, "totalPages": 9, "truncated": false, "ocrPages": 0,
  "tables": 5, "pictures": 6, "markdownChars": 52489, "chunkCount": 42,
  "markdown": "## DocLayNet: A Large Human-Annotated ...",
  "markdownUrl": null,
  "json": null, "jsonUrl": "https://api.apify.com/v2/key-value-stores/…/records/doc-0000-2206.01062-docling.json",
  "chunks": [{ "chunkIndex": 0, "text": "…", "headings": ["DocLayNet …"], "pageNumbers": [1], "tokenCount": 134 }],
  "billedUnits": { "pageConverted": 9, "pageAccurateTables": 0, "ocrPage": 0, "chunkBatch": 1 },
  "processingSeconds": 30.1, "doclingVersion": "2.143.0"
}

• Outputs up to Max inline content size (2 MB by default) are placed directly in the row. Larger ones go to the run's key-value store, and the row links to them through markdownUrl, jsonUrl and chunksUrl. Docling JSON is typically 10–20× the size of the Markdown.
• The Documents, Markdown and RAG chunks dataset views give you a quick table. The run's SUMMARY record totals the pages, OCR pages and failures.
• status is success, partial_success (some pages failed or the per-document timeout hit) or failed (with an error message).

Pricing (pay per event)

Event	Price	When
Actor start	$0.00125 per GB of memory ($0.005 at the default 4 GB)	Once per run
Page converted	$0.003	Each PDF page or PPTX slide converted from its text layer with the default fast table mode. Formats without pages (DOCX, HTML, XLSX, MD, CSV) count one page per started 3,000 characters of actual text (minimum 1 per document). Markup, table padding and link URLs are not counted.
Page converted (accurate tables)	$0.007	Only if you choose Table structure = Accurate: each PDF or image page processed with the TableFormer accurate model. It replaces the $0.003 page charge. DOCX, PPTX, XLSX and HTML keep the $0.003 price in accurate mode, because their tables are read natively.
OCR page	$0.01	A page where OCR produced at least half of the text: scans, photos and image files, or every page with OCR = Force. It replaces the page charge and is never added to it. This applies in accurate table mode too: an OCR page costs $0.01, not $0.017.
RAG chunks	$0.001 per started batch of 100 chunks, per document	Only when chunking is enabled

Every page is billed exactly once, at one of the three page prices.

• Failed documents are never charged.
• Max pages per document caps both work and cost per file.
• The Actor respects your run's maximum total charge. It converts only as many pages as fit, then stops.
• Examples at the default 4 GB: a 10-page digital PDF costs $0.035 including the start fee, or $0.075 with accurate tables. A 100-page scanned book costs $1.005. A 20-page Word document (about 40,000 characters) costs $0.047.

Speed and memory

These figures come from local benchmarks on one vCPU, which matches the default 4 GB memory setting.

Document type	Seconds per page
Plain-text PDF	about 1.5–2 s
Typical paper or report with figures and tables	about 3 s (4.5 s with accurate tables)
Dense statistical tables	about 8 s (about 18 s with accurate tables)
Scanned page with OCR	about 5 s
DOCX, PPTX or HTML	well under 1 s per document

• Peak memory stayed below 2.5 GB on every test file, so 4 GB is the recommended setting.
• 8 GB gives two vCPUs and converts PDFs about 1.8× faster. The start fee is then $0.01.
• Models load in about 3 s per run.

Privacy and data retention

• Documents fetched from URLs are downloaded to a temporary folder inside the run container and deleted right after conversion. They are never written to datasets or key-value stores.
• Uploaded files are deleted from their key-value store after the run by default (Delete uploaded files after run). Turn that off if you want to reuse them.
• Converted outputs (Markdown, JSON, chunks) stay in the run's dataset and key-value store under your normal Apify retention settings. You can also send them to a named store of your choice.
• No document content is sent to any external service.

You must have the right to process the documents you submit. Converting a document does not give you any rights to its content.

Tips

• If a scanned PDF has a garbage text layer (bad earlier OCR), use OCR = Force.
• If you only need text from born-digital PDFs, use OCR = Off. It is fastest; scanned pages come back empty.
• Use Table structure = Accurate only for dense tables with merged or multi-row headers, such as financial statements and statistical releases. It is 1.5–2× slower and costs $0.007 per page instead of $0.003.
• Skip OCR inside figures is on by default. On digital pages, picture and chart regions are not OCR'd, which makes papers about 30–40% faster. The Markdown is identical. You only lose text that is drawn inside raster figures, such as screenshot labels, and that text only ever appears in the JSON. Turn it off if you need that text. Scanned pages are always OCR'd in full.
• For RAG, enable chunking and Output chunks as separate dataset rows, then connect the dataset to the Pinecone, Qdrant or other vector DB integration.
• Multi-column layouts, footnotes and page headers are handled by Docling's layout model. Repeating page headers and footers are moved out of the Markdown body (they stay in the JSON as furniture).

Limitations

• Very complex tables (many merged header rows, dense statistical releases) can occasionally have shifted or merged cells. Check json if you need exact cell coordinates.
• OCR is tuned for Latin and Chinese scripts (PP-OCRv6). Handwriting is not supported.
• Formulas come out as text, and pictures are not described. No vision-language models are used.

Open-source licenses

This Actor's code is MIT. The bundled components are listed below; all allow commercial use, and the full list is in LICENSES.md.

Component	License
Docling	MIT
docling-layout-heron model	Apache-2.0
TableFormer weights	CDLA-Permissive-2.0
RapidOCR and PP-OCR weights	Apache-2.0
all-MiniLM-L6-v2 tokenizer	Apache-2.0
PyTorch	BSD-3
ONNX Runtime	MIT

Support

Questions or issues: open an issue on this Actor's Issues tab, or email Tallow Lane Labs at wdeornellas+tallowlanelabs@gmail.com.