Document to Markdown for RAG and AI Agents

Convert PDF, DOCX, PPTX, XLSX and HTML documents into clean Markdown for retrieval augmented generation, document search and AI agents. This Actor extracts text with deterministic document parsers. It requires no browser, language model, API key, OCR engine or model download. Its narrow parser stack avoids the model and rendering dependencies of heavier document pipelines.

Semantic headings and tables are preserved where the source provides them. Word heading styles become Markdown headings, presentation titles become headings, and each populated workbook sheet becomes a heading followed by a table. HTML scripts, styles, navigation, footers and images are removed. PDF headings are inferred conservatively from whole lines and relative font sizes. PDF titles use meaningful metadata, then the largest useful first-page text, excluding license notices. Font-relative spacing preserves words in tightly spaced PDFs, including LaTeX documents.

PDF pages are normalized when at least 80% of their nonblank characters share a 90°, 180° or 270° text orientation. This also handles page-level /Rotate values. Text and ruled tables are then extracted in that reading frame; other rotated text is excluded. The conservative 80% threshold keeps diagram axis labels from changing the page orientation. On mixed pages below that threshold, horizontal text is kept and rotated text is omitted. Arbitrary angles and pages without a consistent orientation can lose meaningful rotated content.

Useful ruled PDF tables become Markdown tables and are appended after the prose on their page. Very sparse grids, grids with mostly short labels, and tables over 20 columns stay as prose; this can also affect compact numeric tables. Diagram fragments are not promoted to headings based on font size alone. Complex layouts, borderless tables, merged cells and multi-column reading order may need manual review. No parser can guarantee identical visual layout in Markdown.

Input

The default input is a bundled, public-domain HTML example. It runs without network access:

{"documents": [{"example": true}]}

Use a public direct-download URL or a binary file that you placed in an Apify key-value store:

{
  "documents": [
    {"url": "https://example.org/report.pdf"},
    {"key": "REPORT", "storeId": "your-store-id", "filename": "report.docx"}
  ],
  "includeMetadata": true,
  "maxFileSizeMB": 10,
  "maxPages": 100,
  "maxSheets": 20,
  "maxChars": 1000000,
  "chunkSize": 2000
}

Replace the illustrative URL and store identifier with your own sources. Provide exactly one of url, key or example per document. Omit storeId to use the run's default store. Upload Office/PDF records as binary content, or HTML as text; JSON/base64 record bodies are not supported. Set filename for extensionless records. Recognized MIME types and file signatures also identify extensionless downloads. HTTP redirects are followed with DNS and connected-address checks at each hop. Only public HTTP(S) addresses on ports 80/443 are accepted, with no embedded username or password. No linked pages, images or external Office relationships are downloaded.

includeMetadata adds filename, format and byte size. chunkSize=0 disables chunks. A positive value returns consecutive character slices with exclusive endChar offsets, no overlap and no additional charge. Slices can split a table, heading or word; retain the full Markdown for structural processing.

Output

One dataset item is written for every attempted document, including failures. Sources skipped after the 55-second processing budget expires receive a free row with error: "Skipped: run processing budget exhausted". Sources beyond your spending limit are not attempted and have no row. A default run returns:

{
  "url": "example://sample.html",
  "title": "RAG document example",
  "markdown": "# RAG document example\n\nA local example that requires no network access.\n\n## Revenue\n\n| Quarter | USD |\n| --- | --- |\n| Q1 | 120 |\n| Q2 | 180 |",
  "pageCount": null,
  "sheetCount": null,
  "charCount": 143,
  "error": null,
  "metadata": {"filename": "sample.html", "format": "html", "sizeBytes": 372}
}

pageCount counts PDF pages or PPTX slides. DOCX pagination requires a renderer and is returned as null; HTML and XLSX also have no page count. sheetCount is only populated for XLSX. Workbook formulas use their saved cached values; this Actor does not recalculate formulas. The first populated row of each table becomes its Markdown header. Optional chunks contain index, startChar, endChar and text.

Pricing and limits

The price is USD 0.004 per successful document, or USD 4 per 1,000 documents, with no start fee. Failed downloads, empty output, corrupt documents, rejected limits and sources skipped after the processing budget generate no conversion charge. Metadata and chunks are included. A result must be delivered before it becomes eligible for a conversion charge. The Actor stops when your spending limit is reached. A billing failure can leave an already delivered result uncharged and stops further processing.

There are at most 20 sources, 20 MiB per source, 200 PDF pages/slides, 50 sheets and 2,000,000 output characters. Defaults are lower; see the input schema. Office ZIPs are capped at 50 MiB expanded and 5,000 parts; workbooks are capped at 100,000 cells. Limits reject the whole document rather than returning partial output. HTTP operations use a 10-second timeout and up to two retries with exponential backoff for transient errors. A document has a 20-second processing deadline; the run has a 55-second processing budget; remaining sources receive free skipped rows. Split larger batches across runs if needed. Storage and SDK cleanup can extend total wall time. Memory allocation is 512 MiB.

Scanned/image-only PDFs need OCR and return a free error. Encrypted files, invalid archives, unsupported formats, blocked downloads and unavailable KVS records also return clear errors. A malformed top-level input produces one free diagnostic row. Unexpected parser details are hidden to avoid leaking document contents into logs.

Responsible use and retention

Use only public sources or files you have permission to process. Do not submit private third-party documents or unnecessary personal information. Temporary conversion files are removed when processing ends, including failed or timed-out conversions. Uploaded KVS records remain under your control, and derived Markdown remains in the run's dataset. Manage source and dataset retention in your Apify account. The bundled example is dedicated to the public domain (CC0).