PDF Extractor — PDF → Clean Markdown + JSON for LLM & RAG

Give it a PDF URL. Get back clean, LLM-ready Markdown and structured JSON. Title, metadata, per-page text, page count, word count, and an accurate token count — every document, one tidy result. Built for RAG pipelines, AI agents, and anyone who needs the text of a PDF without wrestling with binary parsing, broken encodings, or page-layout noise.

Feed your PDFs to your LLM the way it wants to be fed: as clean Markdown, chunkable by page, with the token budget already counted.

---

✨ What it does

• Robust PDF text extraction — powered by Mozilla's pdf.js (via unpdf), the same battle-tested engine that renders PDFs in Firefox. Handles multi-page documents, embedded fonts, and complex layouts.
• Clean Markdown — normalized whitespace, collapsed noise, optional per-page separators (--- Page N ---) so you can chunk on natural document boundaries.
• Structured JSON — title, metadata (author, subject, keywords, creator, producer, creation/modification dates, PDF version), pageCount, pages[] (per-page text), wordCount, tokenCount, fileSizeBytes, fetchedAt.
• Accurate token counts — counted with the GPT/cl100k-family tokenizer so you know exactly how much context each PDF costs before you send it to a model.
• Token budgeting — optional maxTokens truncates output to fit your context window.
• Page-level chunking — the pages[] array gives you clean text per page, ideal for granular RAG retrieval and citations.
• Robust by design — one bad URL never kills the run. Failed downloads/parses return a clean error record (and are never charged).
• Safe & efficient — file-size guard, %PDF- magic-byte validation, and content-type checks reject non-PDFs before wasting compute.

---

🎯 Use cases

You want to…	This Actor gives you…
Build a RAG knowledge base from PDFs	Clean Markdown + per-page text with token counts, ready to embed.
Ingest research papers (arXiv, journals)	Structured text + metadata your pipeline can index and cite.
Feed reports / whitepapers to an LLM	Pre-counted tokens so you never blow the context window.
Give an AI agent document context	Structured JSON your agent can reason over — no binary noise.
Convert PDF docs to Markdown	Portable, diff-friendly Markdown you can drop into a repo or wiki.
Chunk documents by page for citations	A pages[] array with clean text and page numbers.

---

📥 Input

Field	Type	Default	Description
urls	string[]	— (required)	One or more direct links to PDF files. Each successful PDF is one result.
outputFormat	both | markdown | json	both	Include Markdown, the JSON fields, or both.
includePageText	boolean	true	Include a pages[] array with per-page text in the JSON output.
pageSeparators	boolean	true	Insert --- Page N --- markers between pages in the Markdown.
maxPagesPerRun	integer	1000	Safety cap on PDF URLs processed per run.
maxTokens	integer	0	Truncate Markdown to ~N tokens (0 = no limit).
maxFileSizeMb	integer	100	Skip PDFs larger than this size (recorded as failures, never charged).

Example input

{
  "urls": [
    "https://arxiv.org/pdf/1706.03762",
    "https://www.w3.org/WAI/ER/tests/xhtml/testfiles/resources/pdf/dummy.pdf"
  ],
  "outputFormat": "both",
  "includePageText": true,
  "pageSeparators": true,
  "maxTokens": 0
}

---

📤 Output

One dataset item per PDF. Successful example (truncated):

{
  "url": "https://arxiv.org/pdf/1706.03762",
  "finalUrl": "https://arxiv.org/pdf/1706.03762",
  "statusCode": 200,
  "title": "Attention Is All You Need",
  "metadata": {
    "title": "Attention Is All You Need",
    "author": "Ashish Vaswani et al.",
    "creator": "LaTeX with hyperref",
    "producer": "pdfTeX",
    "creationDate": "D:20170612...",
    "pdfVersion": "1.5"
  },
  "pageCount": 15,
  "fileSizeBytes": 2215244,
  "wordCount": 8823,
  "tokenCount": 13140,
  "fetchedAt": "2026-07-19T02:17:19.954Z",
  "markdown": "Attention Is All You Need\n\n...\n\n---\n\n**Page 2**\n\n...",
  "pages": [
    { "page": 1, "text": "Attention Is All You Need\n\n..." }
  ]
}

Failed URL (returned to a separate failures dataset, not charged):

{
  "url": "https://not-a-real-domain-xyz.com/file.pdf",
  "finalUrl": "https://not-a-real-domain-xyz.com/file.pdf",
  "statusCode": null,
  "error": "getaddrinfo ENOTFOUND not-a-real-domain-xyz.com",
  "fetchedAt": "2026-07-19T02:17:19.831Z"
}

---

🔌 Call it from code (Apify API)

curl -X POST "https://api.apify.com/v2/acts/YOUR_USERNAME~pdf-extractor/run-sync-get-dataset-items?token=YOUR_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"urls":["https://arxiv.org/pdf/1706.03762"]}'

import { ApifyClient } from 'apify-client';

const client = new ApifyClient({ token: 'YOUR_TOKEN' });
const { defaultDatasetId } = await client
    .actor('YOUR_USERNAME/pdf-extractor')
    .call({ urls: ['https://arxiv.org/pdf/1706.03762'] });
const { items } = await client.dataset(defaultDatasetId).listItems();
console.log(items[0].markdown);

---

💸 Pricing (Pay-Per-Event)

Event	Price
Actor start	$0.05 per run
Extracted PDF	$0.003 per successfully extracted PDF

• You only pay for PDFs that succeed — failed URLs are never charged.
• 🎁 Free tier: free-plan users' platform usage is covered by Apify, so you can try it and run small jobs at no cost before scaling up.
• Structured per-page JSON + accurate token counts, purpose-built for LLM/RAG ingestion — at a low, predictable price point.

---

⚖️ Acceptable use

This is a general-purpose format-conversion tool: you supply the PDF URLs and are responsible for having the right to download and use the content you submit. The Actor identifies itself with a descriptive User-Agent and only extracts text from documents you point it at. It does not target any single platform's private API and does not harvest personal data as a feature. Please comply with each source's terms of service and applicable law.

---

🧱 Under the hood

Node.js · unpdf (Mozilla pdf.js core, serverless-friendly) · gpt-tokenizer · Apify SDK. Pure-JS extraction — no headless browser, no system libraries. Stateless — nothing is stored between runs.