ParseCraft¶
Document intelligence: align existing converters, parsers, and OCR/VLM models behind one workflow and one typed output format
ParseCraft is a thin aggregation and alignment layer for document intelligence. Capable converters, parsers, and OCR/VLM models already exist, but each emits a different shape — and many ship demo-grade wrapper code pinned to an outdated Python version and dependency set. ParseCraft keeps every tool isolated behind a backend, translates its output into a single typed intermediate representation (IR), and projects that IR to deterministic Markdown. You get one workflow and one output format, while every heavy runtime stays optional.
The IR is the source of truth. Backends read a document and produce typed pages and chunks; every rendering — Markdown today, HTML/JSON/consumer trees later — is a projection of that same IR, never parsed back into state.
Requirements¶
- Python 3.13 (GIL only — the free-threaded
3.13tbuild is not supported) - Python 3.14 or newer (free-threaded
3.14tbuilds are suspended, not tested, until the heavy runtimes we use ship free-threaded macOS/Windows wheels — see the ADR-0001 amendment) - uv (recommended) or pip to install the package
Supported inputs¶
Plain text, Markdown, and HTML convert on the base install. PDF, images, and the
office family — DOCX, PPTX, XLSX, ODT, RTF, EPUB — route through backends that
must be present, and a .docx needs parsecraft[docling] (the layout-, table-,
and reading-order-aware Word path), parsecraft[pymupdf] (AGPL), or a pandoc
binary on the host. No base install converts an office document.
Per-backend formats, licences, and platform limits: the backend catalog.
OCR/VLM models¶
Every model runs behind its backend inside the ocr, surya, or mineru
extra — no model weights load at import time. Weights download into the managed
asset cache on first use (parsecraft models list shows them). Surya weights
are OpenRAIL-M (restricted use: research, personal, and capped startups) and
need --accept-license on install — check before commercial deployment.
| Backend | Model | Weights licence | Extra | VRAM |
|---|---|---|---|---|
ocr-ovis |
OvisOCR2 (ATH-MaaS/OvisOCR2) |
Apache-2.0 | ocr |
1.0 GB |
ocr-tele |
TeleOCR (StarDoc-AI/TeleOCR, zh + en) |
Apache-2.0 | ocr |
1.2 GB |
ocr-unlimited |
Unlimited-OCR (baidu/Unlimited-OCR) |
MIT | ocr |
3.0 GB |
ocr-qianfan |
Qianfan-OCR (baidu/Qianfan-OCR) |
Apache-2.0 | ocr |
4.0 GB |
ocr-surya |
Surya-OCR 2.5 (datalab-to/surya-ocr-2.5) |
OpenRAIL-M (weights; code Apache-2.0) | surya |
1.0 GB |
ocr-tesseract |
Tesseract engine (OS binary, eng + deu) |
Apache-2.0 (tessdata) | base install | CPU only |
mineru |
MinerU2.5-Pro (opendatalab/MinerU2.5-Pro-2605-1.2B) |
Apache-2.0 | mineru |
4.0 GB (GPU optional) |
Install and Quick Start¶
Install the released package and list the available backends:
uv add parsecraft # or: pip install parsecraft
uv run parsecraft backends
First release: 2026.9.2. The current version is on
PyPI.
Building from a source checkout, running the test suite, and working on the codebase are covered in the development setup.
CLI¶
| Command | Description |
|---|---|
parsecraft backends |
List registered backends; load warnings go to stderr |
parsecraft backends --json |
Emit backend descriptors as JSON |
parsecraft config check |
Validate the effective configuration |
parsecraft config show |
Show the resolved configuration with provenance |
parsecraft --version (-v) |
Print the package version |
uv run parsecraft backends
# No backends registered. (no third-party backend installed yet)
See the CLI reference for every command, flag, exit code, and environment variable.
Configuration¶
Settings live in a layered TOML config: ./parsecraft.toml (project) over
~/.config/parsecraft/config.toml (global), with PARSECRAFT_CONFIG_*
environment variables and CLI flags on top. Secrets are references, never
literals — write api_key = "${MY_KEY}".
[routing]
preference = "quality" # speed | balanced | quality
preferred_backend = "docling" # preference, not a permission
max_passes = 2 # fallback-pass budget
[routing.backend_options.docling]
do_ocr = false # forwarded verbatim to the backend
parsecraft config check validates the effective configuration (unknown
backend or engine names fail loudly); parsecraft config show prints every
effective key with the file or layer it came from. See the
configuration reference
for every key.
How it works¶
source document → backend (analyze / convert) → typed IR → Markdown projection
- Backends are the extension point. A third party ships a
BackendFactoryunder theparsecraft.backendsentry-point group; the core package does not change to add one. - One typed IR.
DocumentResult,PageResult, andStructuredChunkmove between layers, and Markdown is a deterministic, dependency-free projection. - Heavy runtimes stay optional. Model weights, CUDA, and OCR/VLM stacks load only inside a backend factory, so importing ParseCraft stays offline-clean.
- Failures are typed. A failed pass records a
PassFailurerather than silently dropping a page.
Continue with the architecture overview, the IR model, the backend catalog, and the backend authoring guide.
Documentation¶
- Installation
- Quickstart
- Architecture
- Backends
- Write a backend
- Backend catalog
- CLI reference
- API reference
- Development setup
License¶
MIT — see LICENSE for details.