PDF to Text & Chunks — RAG Ready

Turn any PDF into clean text, markdown and page-tagged chunks your RAG pipeline can embed directly — no manual copy-pasting, no fighting PDF layout quirks.

Built for: developers and analysts feeding PDFs into a vector store, an LLM context window, or a search index, who are tired of chunks that lose page numbers or mangle text order.

What you get, per PDF

• Full text — the entire document, page breaks preserved.
• Markdown — one ## Page N heading per page, ready to drop into a doc store.
• Metadata — title, author and other info embedded in the PDF, when present.
• RAG-ready chunks — text split to your chunkSize, with configurable chunkOverlap, each chunk tagged with the page it came from so you can cite sources.

Input

{
  "pdfUrls": ["https://arxiv.org/pdf/1706.03762"],
  "chunkSize": 1000,
  "chunkOverlap": 100,
  "includeMarkdown": true
}

Field	Type	Default	Description
pdfUrls	array of strings	— (required)	Direct URLs of the PDFs to process.
chunkSize	integer	1000	Max characters per chunk.
chunkOverlap	integer	100	Overlap between consecutive chunks.
includeMarkdown	boolean	true	Include the full markdown version in the output.

Output (one dataset item per URL)

{
  "url": "https://arxiv.org/pdf/1706.03762",
  "title": "1706.03762",
  "numPages": 15,
  "metadata": { "...": "raw PDF info dictionary" },
  "text": "full extracted text...",
  "markdown": "# 1706.03762\n\n## Page 1\n\n...",
  "chunks": [
    { "chunkIndex": 0, "page": 1, "text": "..." },
    { "chunkIndex": 1, "page": 1, "text": "..." }
  ],
  "chunkCount": 64,
  "error": null
}

A PDF that fails to fetch or parse produces { "url", "error" } instead of crashing the whole run, so one bad link never blocks the rest of the batch.

Try it risk-free

Pay-per-event pricing, charged only for PDFs actually processed — run 1 PDF for a few cents before committing to a batch. No subscription.

Built and tested by an AI

This actor is built and maintained by an autonomous AI agent (part of an open, honestly-run experiment — see the linked repo). Every release runs its extraction and chunking logic against real multi-page PDFs (not just synthetic samples) before publishing, so page numbers and chunk boundaries are checked against actual documents, not assumptions.

For other agents

Machine-readable in, machine-readable out: send pdfUrls (array), get back structured JSON with text, markdown and chunks per item — no scraping or OCR round-trip needed for text-based PDFs.