PDF to Markdown & JSON Converter (Docling)

Convert PDF documents into clean Markdown, structured JSON, and plain text using IBM's open-source Docling AI library. Handles regular text-based PDFs and scanned/image PDFs (with OCR), extracts tables with high accuracy, and optionally saves embedded images.

No external API key required. Docling runs entirely on-device inside the actor — nothing leaves the platform.

---

Features

• 7 output formats in one run: Markdown (great for RAG pipelines), structured JSON (full document hierarchy), and plain text
• Table extraction: Docling's TableFormer model detects and exports tables with accurate row/column structure — available as Markdown tables in the Markdown output and as structured arrays in JSON
• Optional OCR: Enable for scanned PDFs or image-based pages (using EasyOCR, Apache 2.0 licensed — no GPL dependencies)
• Image extraction: Save embedded images as PNG files to the key-value store
• Batch processing: Supply multiple URLs — each document is converted independently with per-item error isolation
• Open protocol: Official IBM Research library, MIT licensed, no ToS grey area

---

Use cases

• RAG / AI pipelines — convert a library of PDFs to Markdown for ingestion into vector databases (LangChain, LlamaIndex, Haystack)
• AI training data — extract clean text and structured JSON from research papers, reports, and books
• Document digitisation — make scanned archives searchable with OCR
• Data extraction — extract structured tables from financial reports, scientific papers, legal documents
• Content migration — convert PDF documentation to Markdown for static-site generators or wikis

---

Input

Field	Type	Default	Description
pdfUrls	string[]	(required)	HTTP/HTTPS URLs to PDF files
outputFormats	string[]	["markdown","json"]	Which formats to produce: markdown, json, text
enableOcr	boolean	false	Run OCR on each page (needed for scanned PDFs)
tableMode	string	"FAST"	Table extraction: FAST or ACCURATE (complex tables)
extractImages	boolean	false	Save embedded images as PNG to the key-value store
proxyConfiguration	object	—	Optional proxy for PDF downloads

Example inputs

Basic — single PDF to Markdown + JSON:

{
  "pdfUrls": ["https://arxiv.org/pdf/1706.03762"],
  "outputFormats": ["markdown", "json"]
}

Scanned PDF with OCR + accurate table extraction:

{
  "pdfUrls": ["https://example.com/scanned-report.pdf"],
  "outputFormats": ["markdown", "text"],
  "enableOcr": true,
  "tableMode": "ACCURATE"
}

Batch with image extraction:

{
  "pdfUrls": [
    "https://example.com/report-q1.pdf",
    "https://example.com/report-q2.pdf"
  ],
  "outputFormats": ["markdown", "json"],
  "extractImages": true
}

---

Output

Dataset record (one per document)

{
  "url": "https://arxiv.org/pdf/1706.03762",
  "filename": "1706.03762.pdf",
  "status": "success",
  "pageCount": 15,
  "wordCount": 8241,
  "tableCount": 4,
  "imageCount": 0,
  "markdownKey": "1706_03762_a1b2c3d4.md",
  "jsonKey": "1706_03762_a1b2c3d4.json",
  "textKey": null,
  "markdownUrl": "https://api.apify.com/v2/key-value-stores/.../records/1706_03762_a1b2c3d4.md",
  "jsonUrl": "https://api.apify.com/v2/key-value-stores/.../records/1706_03762_a1b2c3d4.json",
  "textUrl": null,
  "errorMessage": null,
  "scrapedAt": "2026-05-24T12:00:00.000Z"
}

status values: success, partial (some pages failed), error (full failure; see errorMessage).

Key-value store

For each successfully converted document, the actor writes:

• {stem}_{hash}.md — Markdown with tables rendered as GFM tables
• {stem}_{hash}.json — Full Docling document JSON (hierarchy, bounding boxes, cell-level data)
• {stem}_{hash}.txt — Plain text (if text format selected)
• {stem}_{hash}_img_N.png — Embedded images (if extractImages is true)

---

Pricing

Pay per event — you only pay for what you actually process.

Event	Price	When charged
Page processed	$0.005	Per page converted (no OCR)
OCR page processed	$0.020	Per page converted with OCR
Table extracted	$0.002	Per table detected
Image extracted	$0.001	Per image saved

Example costs:

• 100-page report, no OCR, 10 tables: 100 × $0.005 + 10 × $0.002 = $0.52
• 50-page scanned report with OCR, 5 tables: 50 × $0.020 + 5 × $0.002 = $1.01
• 1,000 research papers (avg 10 pages, 3 tables, no OCR): ~$58

---

Throughput & memory

• Memory: 4 GB minimum (PyTorch + layout models). Increase to 8 GB for large PDFs or ACCURATE table mode.
• Speed (no OCR): ~5–10 pages/min on a single CPU core.
• Speed (OCR): ~1–3 pages/min on a single CPU core.
• Cold start: ~10–30 seconds for model initialisation (models are pre-baked into the Docker image; no download at runtime).
• For large batches, use multiple actor runs in parallel — Apify handles scheduling automatically.

---

Authentication

No authentication required. Docling accesses no external APIs — all processing is on-device.

The optional proxyConfiguration field is only used for downloading PDFs from URLs that may require a proxy (e.g., behind a corporate firewall). Conversion itself is always local.

---

FAQ

Q: Does it work on scanned PDFs? A: Yes — enable enableOcr: true. This runs EasyOCR (Apache 2.0) on each page. Expect ~1–3 pages/min and higher per-page cost.

Q: How accurate is table extraction? A: Docling's TableFormer model is state-of-the-art for PDF table recognition. Use tableMode: "ACCURATE" for complex multi-column or merged-cell tables. Simple tables work well in FAST mode.

Q: What's in the JSON output? A: The full Docling DoclingDocument serialised to JSON — includes document hierarchy (sections, paragraphs, tables, figures), bounding boxes, page references, and cell-level table data. Ideal for downstream parsing.

Q: Can I process password-protected PDFs? A: No. Password-protected PDFs cannot be opened by Docling. Remove the password before uploading.

Q: How does it compare to PDF.js or pdfplumber? A: Docling uses AI layout analysis and TableFormer for structural understanding — it correctly identifies headings, columns, captions, and complex tables in ways that regex/heuristic tools cannot. The tradeoff is higher memory and CPU cost.

---

Legal

Docling is released under the MIT License by IBM Research. The OCR backend (EasyOCR) is Apache 2.0. No GPL-licensed components are included.

This actor downloads and processes PDFs you supply. You are responsible for ensuring you have the right to process the content of the PDFs under applicable copyright, data-protection (GDPR/CCPA), and terms-of-service law. Do not use this actor to process content you do not have permission to reproduce or analyse.