anytomd

Python binding of anytomd-rs — the pure-Rust engine by developer0hye that converts DOCX, PPTX, XLSX, HTML, CSV, notebooks, images and code into clean Markdown, designed for LLM consumption.

Same engine, same output, now one pip install away: no Python dependencies, the GIL released during conversion, native asyncio, typed end to end.

pip install anytomd        # or: uv add anytomd

from anytomd import convert_file

result = convert_file("report.docx")
print(result.markdown)     # headings, tables, lists, links, images
print(result.plain_text)   # the same content, no Markdown syntax
print(result.title)        # 'Quarterly Report'

Formats

Detected from magic bytes and extension. PDF is refused by the engine on purpose: current LLMs read it natively — the optional pdf extra parses it locally when you would rather not send it. Details per format → Formats.

API

Options, keyword-only on every conversion: extract_images, extract_comments, strict, max_input_bytes, max_total_image_bytes, max_uncompressed_zip_bytes, and image_describer / async_image_describer. One by one → Reference.

LLM image descriptions

Any callable (image: bytes, mime_type: str, prompt: str) -> str writes the alt text — your OpenAI, Anthropic or local model — or use the built-in Gemini describers:

from anytomd import GeminiDescriber, convert_file

result = convert_file("deck.pptx", image_describer=GeminiDescriber())   # GEMINI_API_KEY
# ![A bar chart of quarterly revenue by region](image1.png)

The async variants take an async def and describe every image of a document concurrently. A failing describer never aborts the conversion: it becomes a warning. → Image descriptions

Optional capabilities

The wheel keeps zero Python dependencies. Four extras add a capability — the name says what it brings, not the library behind it:

GeminiDescriber needs no extra: it is compiled into every wheel.

ImageDescriber composes them with one rule — empty means no letters: OCR answers images that carry text; only when it reads nothing does the vision model describe the image; when everything is silent, the alt text stays empty exactly as without a describer.

uv add "anytomd[full]"

from anytomd import ImageDescriber, convert_file

describer = ImageDescriber(
    model="claude-sonnet-5-5",
    api_base="https://api.anthropic.com/v1/",       # Anthropic's OpenAI-compatible endpoint
    api_key="sk-ant-...",
)
result = convert_file("deck.pptx", image_describer=describer)
describer.close()

Errors

Best-effort by default: recoverable problems land in result.warnings; strict=True raises instead. Unrecoverable ones raise a subclass of ConvertError; unreadable paths raise FileNotFoundError and friends. → Errors and warnings

Upstream

All conversion logic, format coverage and performance belong to anytomd-rs — see its README for benchmarks, the CLI and the WASM build. anytomd.__version__ reports the engine release compiled into your wheel. Markdown output issues go upstream; Python API issues go here. → Upstream and versioning

Self-maintenance

The binding follows its dependencies on its own; a human only reviews and merges.

Package version = engine version; binding-only fixes ship as .postN. Exactly which files each piece may edit, and with which tool → What it touches

Development

uv sync && uv run prek install   # hooks: ruff, ty, tach, rustfmt, clippy, licenses, zizmor, docs, gitleaks
make check                       # build + ruff + ty + tach + tests
make cov                         # Rust coverage driven by the Python suite (gate 90 %)

→ Development

License

Apache-2.0, like anytomd-rs. Wheels statically link anytomd-rs and its dependencies; their licenses ship in THIRD_PARTY_LICENSES.md, inside every wheel too.