AILANG Parse

Universal document parsing and generation in AILANG. Extracts structured content from DOCX, PPTX, XLSX, PDF, and image files into JSON and markdown — and writes documents back out in 9 formats. To author a document, write Markdown and convert it; see Writing documents in Markdown.

Office formats (DOCX, PPTX, XLSX) use deterministic XML parsing — no AI, no cloud, instant results. PDFs default to the deterministic pdftotext backend (poppler) — also no AI, no cloud — with docling and liteparse as local alternatives escalating automatically to local docling OCR for scanned/image-only PDFs. Pluggable AI (Gemini, Claude, local Ollama) is available via --pdf-backend ai but is never chosen automatically. Images delegate to whatever AI model you plug in. AILANG Parse is AI-agnostic: swap --pdf-backend/--ai to change the backend, zero code changes.

Install

Two routes; both put a docparse command on your PATH, and both write output next to the file you parse (--output-dir DIR or DOCPARSE_OUTPUT_DIR to change that).

Already have AILANG? (0.42.0+ of this package, AILANG v0.40.0 dev or later)

ailang install sunholo/ailang_parse
docparse report.docx

The package's [bin] table gives you a shim in ~/.ailang/bin that runs straight out of the registry cache: parse one file, or --convert it. It is deliberately the thin path — describe/summarize, batch mode, PDF backends and --install-backends are not on it (see below). ailang bin list / ailang bin uninstall docparse manage it.

Otherwise, or for the full CLI:

curl -fsSL https://www.sunholo.com/ailang-parse/install.sh | sh

Fetches the published package (~400 KB), installs the AILANG runtime if you do not have it, and links the full docparse wrapper (plus docparse-audit and docparse-render) into ~/.local/bin. --version, --prefix and --uninstall are supported; re-running is a no-op. If you install both, ailang install tells you which one docparse resolves to.

That covers every deterministic format. PDF needs two more things, and they are easy to miss:

brew install poppler        # pdftotext, the default PDF backend
                            # (apt install poppler-utils on Debian/Ubuntu)
docparse --install-backends # docling + liteparse, for scans and layout

--install-backends matters even if you never pass --pdf-backend: when pdftotext finds no text layer the parser escalates to docling on its own, so without it a scanned PDF fails on the default backend. AI backends authenticate with Google ADC (gcloud auth application-default login), not an API key.

git clone https://github.com/sunholo-data/ailang-parse.git
ln -s "$(pwd)/ailang-parse/bin/docparse" /usr/local/bin/docparse

The wrapper finds the project root by walking up for docparse/main.ail, so it works from a clone, an installed prefix, or a symlink chain. Note that the CLI wrapper and the PDF adapter live under assets/ — the only path the AILANG publisher bundles verbatim — with symlinks at their historical locations.

SDKs

Use AILANG Parse from your language of choice:

pip install ailang-parse          # Python
npm install @ailang/parse         # JavaScript/TypeScript
go get github.com/sunholo-data/ailang-parse-go  # Go

Quick Start

# Office documents (deterministic, no AI needed)
docparse report.docx
docparse slides.pptx
docparse spreadsheet.xlsx

# PDF (deterministic pdftotext by default — no AI); images (AI auto-enabled)
docparse document.pdf
docparse photo.png

# Options
docparse report.docx describe        # AI image descriptions
docparse report.docx summarize       # AI document summary
docparse contract.pdf                # PDF: deterministic pdftotext (default)
docparse scan.pdf                    # Scanned PDF: escalates to local docling OCR
docparse scan.pdf --pdf-backend ai --ai gemini-2.5-flash  # External AI (paid), opt-in only

# Format conversion
docparse report.docx --convert output.html
docparse data.csv --convert report.docx
docparse notes.md --convert slides.pptx
docparse notes.md --convert offer.docx --reference-doc letterhead.docx

# AI document generation
ailang run --entry main --caps IO,FS,Env,AI --ai gemini-2.5-flash \
  docparse/main.ail --generate report.docx --prompt "Q1 sales report with tables"

Output

Every run produces, in the directory you ran from (--output-dir DIR or DOCPARSE_OUTPUT_DIR to change it):

• <filename>.json — Structured JSON with typed blocks
• <filename>.md — LLM-ready markdown

What AILANG Parse Extracts

Feature	DOCX	PPTX	XLSX	Best Competitor
Tables with merged cells	Yes	Yes	Yes	Raw OOXML only
Track changes (redlining)	Yes	—	—	Pandoc (3/3)
Comments (interleaved)	Yes	—	—	Raw OOXML (2/2)
Headers/footers	Yes	—	—	Kreuzberg (2/3)
Text boxes / VML shapes	Yes	Yes	—	Raw OOXML (1/2)
Equations (§22.1)	Yes	—	—	None
Field codes (§17.16)	Yes	—	—	Kreuzberg, OOXML
Speaker notes	—	Yes	—	None
Multi-sheet extraction	—	—	Yes	Kreuzberg

OfficeDocBench (69 files, 11 formats, 7 metrics): AILANG Parse 93.9% composite with 100% coverage vs nearest competitor 68.0% coverage-adjusted. 8 parsers compared including Raw OOXML, Pandoc, Kreuzberg, MarkItDown, Unstructured, Docling. Scores include aspirational ECMA-376 spec targets that intentionally lower our score.

Supported Formats

Parsing (16 formats): DOCX, PPTX, XLSX, ODT, ODP, ODS, HTML, Markdown, CSV, EPUB, EML, MBOX, TEX, RTF, PDF, images (JPG/PNG)

Generation (9 formats): DOCX, PPTX, XLSX, ODT, ODP, ODS, HTML, Markdown, QMD (Quarto)

Writing documents in Markdown

Markdown is the input an LLM can write, so it is the practical way to generate a document: write markdown, convert to any of the nine output formats.

docparse report.md --convert report.docx

What survives the trip: YAML front matter (title/author/date → document properties), bold/italic/code/strike as real character formatting, links as real hyperlinks, images (local paths are read and embedded), fenced code blocks, blockquotes, nested lists, thematic breaks, and tables with alignment and column spans.

Headers, footers, comments and tracked changes have no Markdown syntax; those are preserved when converting from a document that already contains them.

Local document verification (v0.41.2+)

docparse-audit and docparse-render are AILANG companions to the CLI. AILANG Parse v0.41.2 requires AILANG 0.35.0 or newer. They run locally and require no hosted account. The installer links them when present in the installed release; in a checkout use ./bin/docparse-audit and ./bin/docparse-render.

docparse-audit report.docx --strict
docparse-render report.docx --output-dir qa/report
docparse-render before.docx --compare after.docx --output-dir qa/comparison

The read-only DOCX audit emits JSON inventories and findings for table grids and headers, heading gaps, image descriptions, links, comments and fields. Exit codes are 0 for a completed audit, 1 for findings with --strict, and 2 for unreadable input. It is a bounded structural check; it does not certify accessibility or resolve arbitrary style inheritance.

Rendering supports .docx, .odt, .pptx, .xlsx and .pdf. Install LibreOffice and Poppler for Office input; PDF needs only Poppler. No new Python dependency is needed. Select tools with --soffice / --pdftoppm (or DOCPARSE_SOFFICE / DOCPARSE_PDFTOPPM). --dpi defaults to 144 and --timeout to 120 seconds per process. The output directory must be new. Each document gets a PDF and numbered PNG pages; the manifest records source hashes, tool paths and page directories. Comparison identifies changed/added/removed pages using uncompressed raster hashes. Inspect the two PNG sets side by side; no diff overlays are generated. visual_review: pending means page images still need inspection.

The implementation is in docparse/services/document_render.ail and docparse/services/docx_audit.ail; regression verification is AILANG too:

ailang run --entry main --caps IO,FS,Env,Process \
  --process-allowlist "true,false,mktemp,mkdir,$(command -v soffice),$(command -v pdftoppm)" \
  --process-timeout 120s scripts/verify_document_quality.ail -- \
  /tmp/docparse-quality-new "$(command -v soffice)" "$(command -v pdftoppm)"

Styling a generated DOCX from a template

--reference-doc is the Quarto/Pandoc reference-doc feature: an existing .docx supplies the look, the Markdown supplies the content.

docparse annex.md --convert annex.docx --reference-doc letterhead.docx

Generated tables use the selected section's usable width, with bounded content weights for columns and repeating semantic headers. Merged-cell widths sum the columns they cover.

The template's styles.xml, numbering.xml, theme, embedded fonts, headers, footers and page setup are applied to the new content. Everything the merge does not regenerate is carried through byte-for-byte, so the letterhead, logo and licensed fonts come out exactly as they went in.

What comes from where:

Template	page size, margins, headers, footers, page numbering, fonts, theme, colours
Your document	the body content, and docProps/core.xml (title/author)
Merged	styles.xml (ours fill only the styleIds the template lacks), numbering.xml (our list definitions take ids above the template's), [Content_Types].xml, both .rels

Two consequences worth knowing:

• The template's headers and footers win. A source document's own headers are dropped rather than mixed with the letterhead. The page furniture all lives in the template's body <w:sectPr>, which is lifted whole.
• The template's comments are dropped along with its body, and so are commentsExtended.xml and people.xml. Comments in the source document still come through.

Two flags refine a multi-section template:

• --reference-section N picks which of the template's sections supplies the page setup, headers and footers — 1 is the first section, Word's numbering. The default is the last section (the body-level one, what the flag-less behaviour has always lifted). A multi-section template's wanted furniture is often an earlier section's — the master agreement's CONFIDENTIAL footer, not the Annex's missing one.
• --table-style NAME binds generated tables to a table style the template defines (matched on styleId, then style name). Without it, the style named Table is used if the template has one, else the first table style that is not the implicit Normal Table. Under a bound style the generator stops emitting its own hardcoded borders — the style carries them.

An unreadable or non-DOCX reference is an error and writes nothing — a silent fallback to the built-in styling would produce a plausible file missing exactly the letterhead it was asked for. DOCX output only.

Architecture

docparse/
├── types/document.ail           # Block ADT (11 variants)
├── services/
│   ├── format_router.ail        # Format detection (36 inline tests)
│   ├── zip_extract.ail          # ZIP layer (9 inline tests)
│   ├── docx_parser.ail          # DOCX XML → Blocks (6 inline tests)
│   ├── pptx_parser.ail          # PPTX slides → Blocks
│   ├── xlsx_parser.ail          # XLSX worksheets → Blocks
│   ├── direct_ai_parser.ail     # PDF/image → Blocks (AI)
│   ├── layout_ai.ail            # AI self-healing (optional)
│   ├── output_formatter.ail     # JSON + markdown output
│   └── docparse_browser.ail     # WASM browser adapter
└── main.ail                     # CLI entry point

91 contracts, 50+ inline tests. Of the 91, Z3 proves 14 outright; the rest are checked at runtime under --prove/--verify-contracts in CI, and skip statically because parser code is recursive and higher-order, which is outside Z3's decidable fragment.

AI Configuration

AILANG Parse uses AILANG's AI effect — any model AILANG supports works:

docparse scan.pdf --ai gemini-2.5-flash          # Google (default; fast)
docparse scan.pdf --ai gemini-3-flash-preview    # Google (slower; thinking model)
docparse scan.pdf --ai granite-docling           # Local Ollama (free)
docparse scan.pdf --ai claude-haiku-4-5          # Anthropic

AI usage is bounded by capability budgets (AI @limit=200 on main), so costs are predictable.

Dev Commands

docparse --check       # Type-check all modules
docparse --test        # Run inline tests
docparse --prove       # Static Z3 contract verification

Benchmarks

uv run benchmarks/run_benchmarks.py --suite office     # Structural (no API, instant)
uv run benchmarks/run_benchmarks.py --suite pdf         # PDF extraction (needs AI)
uv run benchmarks/run_benchmarks.py --competitors       # Compare to Docling etc.

See benchmarks/ for details.

License

Apache 2.0