PDF, Word & PowerPoint to Markdown + RAG Chunks

Turn PDFs, Word documents and PowerPoint decks into clean Markdown that LLMs read well, and into chunks ready for a vector database: token-counted, split along headings, with the heading path and page numbers attached to every chunk.

Use it to load company documents into a RAG chatbot, feed reports to GPT/Claude, build a searchable knowledge base, or migrate documents into Notion or a wiki.

What the Markdown keeps

	PDF	DOCX	PPTX
Headings (#, ##, ...)	from font size and bold (body text size is detected per document)	from Word heading styles + Title	one ## Slide N: title per slide
Tables → Markdown tables	ruled/grid tables (borderless tables stay as text in reading order)	all tables	tables, plus chart data as a table
Lists	bullets and numbering	bullets and numbering	bullet levels
Clean-up	repeated headers/footers and page numbers removed, hyphenated line breaks joined, rotated margin text and figure labels skipped, multi-column reading order	images dropped	speaker notes (optional)

Scanned (image-only) PDF pages are detected and reported in scannedPages. They need OCR, which this Actor doesn't do, and they are not charged.

Chunks built for retrieval

• Token-sized with the OpenAI cl100k tokenizer (chunkSizeTokens, default 500) with optional overlap
• Heading-aware: a new section starts a new chunk, and each chunk is prefixed with its path, e.g. [Annual Report > 3. Risks > Currency]
• Tables are never cut mid-row. Long tables are split between rows and the header row is repeated in every piece.
• pageStart / pageEnd on each chunk for citations ("see page 7")
• Stable chunkId (<documentId>-0003) so re-runs can upsert into Pinecone, Qdrant, Weaviate, pgvector...

Input

{ "documents": [{ "url": "https://arxiv.org/pdf/1706.03762" }, { "url": "https://drive.google.com/file/d/<id>/view" }],
  "chunkSizeTokens": 500, "chunkOverlapTokens": 50, "outputMode": "chunks" }

Google Drive, Google Docs/Slides (exported as DOCX/PPTX) and Dropbox share links are converted to direct downloads. The file must be shared as "anyone with the link".

Output (chunk mode)

{ "fileName": "report.pdf", "documentId": "3ab677dde1ab", "title": "Quarterly Report", "status": "ok",
  "chunkId": "3ab677dde1ab-0002", "chunkIndex": 2, "chunkCount": 7, "tokens": 276,
  "headingPath": "Quarterly Report 2026 > 1. Summary > 1.1 Regional breakdown", "pageStart": 1, "pageEnd": 1,
  "text": "[Quarterly Report 2026 > 1. Summary > 1.1 Regional breakdown]\n\n| Region | Q1 | Q2 | Q3 | Change % |\n|---|---|---|---|---|\n| South 0 | 8,233 | ...",
  "markdownUrl": "https://api.apify.com/v2/key-value-stores/.../records/0001-report.md" }

In documents mode you get one item per file with the whole markdown string. Either way, the full .md file and a .meta.json (pages, tables, scanned pages) are in the key-value store.

Tested on

Our test corpus checks: a Word report and the same report exported to PDF (3 tables, 100% of table cells reproduced exactly in both), a two-column arXiv paper (correct section headings, side stamp removed), the IRS W-9 form, a PPTX with nested bullets, a table and speaker notes, and a scanned PDF (correctly flagged). Bad links, HTML pages instead of files, 404s and password-protected PDFs return clear errors.

Pricing

Pay per event: page-converted, one per PDF page with text, per slide, or per 500 words of DOCX/Markdown. Scanned pages, failed downloads and unsupported files are free. Use maxPagesPerDocument and Max cost per run to cap spend.