Document to Markdown & RAG Chunks — PDF, Word, Excel, PowerPoint, HTML (CJK-ready)

Run it on Apify Store: https://apify.com/fontom.ai.plus/document-to-markdown

Turn documents into clean Markdown, plain text and RAG-ready chunks for LLM agents, search indexes and data pipelines. Pure code running inside the Actor: no OCR service, no AI guessing, no third-party API, no API keys.

What makes the output good

• Two-column PDFs are read in the right order (left column, then right; full-width titles first). Most simple extractors interleave the columns line by line.
• PDF tables come out as real Markdown tables (header row, right-aligned numbers, wrapped cell text, Japanese tables), found from the alignment of the text itself. Turn it off with detectTables: false.
• Headings, bullet lists and wrapped paragraphs in PDFs are rebuilt from font sizes and positions; page-by-page text is included.
• Japanese / Chinese / Korean text works: no stray spaces inside words, and PDFs that encode characters as Kangxi radicals (⽇ ⺠ ⽂ instead of 日 民 文, common in Wikipedia and government PDFs) are folded back to normal ideographs so search and embeddings match. Optional NFKC normalisation for full-width letters and digits.
• Word (headings, lists, tables), Excel (every sheet as a Markdown table), PowerPoint (slide by slide, titles, bullets, speaker notes), HTML (main content only: navigation, headers, footers, forms, scripts and images dropped), CSV / TXT / MD.
• Shift_JIS, EUC-JP and UTF-16 files are decoded correctly (Japanese government CSVs and older sites are often not UTF-8). The detected encoding is reported when it is not UTF-8.
• Scanned PDFs are flagged clearly: scanned: true, the page numbers without a text layer, and a warning, so a pipeline can route them to OCR.
• RAG chunks (optional): paragraph- and sentence-aware, with overlap and the page range of every chunk.
• Safe by design: only public http(s) URLs (private, local and cloud-metadata addresses are blocked, also after redirects), file size and page limits.

Input

Field	Meaning
urls	Public document URLs (up to 5,000 per run).
documents	Optional inline files: [{"name": "a.pdf", "base64": "..."}].
chunkSize	Characters per chunk (0 = off; 1000–2000 typical). chunkOverlap (default 150).
maxPages, maxFileMb, maxChars	Limits (defaults 500 pages, 50 MB, 2,000,000 characters).
includePageTexts	Page-by-page text for PDFs (default on).
detectTables	Turn aligned PDF text into Markdown tables (default on). Switch off to get plain lines.
fixCjkVariants	Fold Kangxi-radical/compatibility glyphs (default on). normalize = NFKC (default off).

Output (one row per document)

{ "ok": true, "source": "https://example.com/report.pdf", "type": "pdf", "title": "…", "pages": 12, "script": "ja",
  "chars": 18234, "markdown": "# …", "text": "…", "pageTexts": ["…"], "chunks": [{ "index": 0, "text": "…", "chars": 1432, "pages": [1, 2] }],
  "warnings": [] }

script is ja, zh, ko, latin, other or unknown, detected from the characters present. tables counts the Markdown tables found in a PDF. scanned / emptyPages appear for PDFs without a text layer, encoding for files that were not UTF-8. warnings tells you when something was skipped.

Pricing

Pay per event, as shown on the Actor page: a tiny start fee per run ($0.00005) plus one charge per result row ($0.01) (Apify's dataset-item event). Every row is charged, including rows for documents that failed to convert (404, blocked address, unsupported type, too large), so check URLs before big batches. One run can hold up to 5,000 documents.

Limits, stated plainly

• No OCR. Scanned or image-only PDFs return a warning ("no text layer"), not text.
• Vertical (tategaki) Japanese reading order is not reconstructed (a warning is raised when it is detected).
• PDF tables are inferred from text alignment, so they are only as good as the layout: borderless tables, merged header cells and tables that continue across pages can be split or mis-aligned, and infographics or figure labels are occasionally turned into a small table. Tables that sit beside body text in a two-column page are read as text. Use detectTables: false if you prefer plain lines.
• Legacy .doc, .xls, .ppt are not supported (save as .docx, .xlsx, .pptx). Password-protected files fail with an error row.
• Layout-heavy PDFs (magazines, posters, forms) can come out in an imperfect reading order. Multi-column detection handles the common two-column paper/report layout only.
• Only documents you are allowed to process: you are responsible for the rights to the files you convert.

Tested on real, public documents (conversion time on a laptop, tables found in the PDF): Attention Is All You Need (15 pages), BERT (16), LLaMA (27), GPT-3 (75, 244k characters), the Japanese National Tax Agency invoice-system leaflets (27 and 4 pages), the Ministry of Internal Affairs and Communications white-paper index tables (8 pages, long Japanese tables spanning pages), a Wikipedia Japanese PDF, the e-Gov law Word files and a Japanese HTML page, plus unit tests for every file type and for tables, text columns, wrapped cells, Japanese tables, Shift_JIS/EUC-JP files and scanned pages. The check script is scripts/benchmark.mjs.