[build-system]
requires = ["hatchling"]
build-backend = "hatchling.build"

[project]
name = "mdcx"
version = "1.31.0"
description = "Convert a document collection to verified Markdown, package it encrypted with its search index and provenance, and serve it to agents over MCP"
readme = "README.md"
requires-python = ">=3.11"
license = "Apache-2.0"
license-files = ["LICENSE", "NOTICE"]
authors = [{ name = "Jorge Ellena G." }]
keywords = ["markdown", "document-conversion", "retrieval", "mcp", "provenance", "rag"]
classifiers = [
    "Development Status :: 4 - Beta",
    "Intended Audience :: Developers",
    "Intended Audience :: Legal Industry",
    "Intended Audience :: Science/Research",
    "Topic :: Text Processing :: Markup :: Markdown",
    "Topic :: Text Processing :: Indexing",
    "Topic :: Office/Business",
    "Topic :: Scientific/Engineering :: Information Analysis",
    "Programming Language :: Python :: 3",
    "Programming Language :: Python :: 3.11",
    "Programming Language :: Python :: 3.12",
    "Programming Language :: Python :: 3.13",
    "Operating System :: OS Independent",
    "Environment :: Console",
    "Typing :: Typed",
]

# Querying a package requires almost nothing: SQLite and LZMA are in the standard
# library, and only decryption needs a third-party module. Conversion is what pulls in
# heavy dependencies, so it is an extra rather than a default: a user who receives an
# .mdcx file and wants to query it should not install 1.4 GB to do so.
dependencies = [
    "cryptography>=41.0",
]

[project.optional-dependencies]
# Document conversion. Docling brings PyTorch and its model stack.
convert = [
    "docling>=2.0",
    "pypdfium2>=4.0",
    "python-docx>=1.0",
    "openpyxl>=3.1",
    "python-pptx>=1.0",
]
# Model Context Protocol server.
mcp = ["mcp>=1.0"]
# Packages compressed with zstd. Opening is what a server pays before it can
# answer anything, and on a large body zstd reads back in a fraction of the
# time LZMA takes. Optional because LZMA is in the standard library and every
# package written before this is compressed with it.
zstd = ["zstandard>=0.22"]
# Retrieval across languages. A query written in one language reaches documents
# written in another, which word matching cannot do: the words are not shared.
# It brings a multilingual embedding model and its runtime.
multilingual = [
    "sentence-transformers>=3.0",
    "numpy>=1.24",
]
# Optical character recognition, for collections containing scanned documents.
#
# `onnxruntime` and `onnxruntime-gpu` are two distributions publishing the same
# module, so they cannot coexist: whichever pip wrote last wins, and in practice
# that is the CPU one. Declaring the CPU build here means `pip install --upgrade
# "mdcx[all]"` silently removes CUDA from an environment that had prepared it --
# measured on three consecutive releases, in two environments each time, with no
# error, nothing in any log, and optical recognition simply costing tens of
# times more.
#
# pip cannot express "either of these distributions", so `ocr-gpu` exists: the
# same extra without the runtime pinned, for an environment that installs
# `onnxruntime-gpu` itself. `mdcx[all-gpu]` is the same idea for the whole set.
#
# `rapidocr-onnxruntime` is gone rather than replaced. It is the package that
# pulled `onnxruntime` in, nothing requires it, mdcx never imports it, and
# docling has moved to `rapidocr` 3.x.
ocr = ["rapidocr>=3.0", "onnxruntime>=1.16"]
ocr-gpu = ["rapidocr>=3.0"]
# Tables a page does not draw. Most tables in printed material are found from
# the rules drawn around them, which costs nothing; what is left is the
# borderless kind, and those are read by a model that reports where the rows
# and columns run. It reads the shape only -- the words still come from the
# text layer -- and it is small: twenty-nine million parameters, run on the few
# pages that reached it. Licences are compatible with distributing a library:
# code and weights MIT, PubTables-1M CDLA-Permissive-2.0.
tables = [
    "transformers>=4.40",
    "timm>=1.0",
    "torch>=2.0",
]
# Everything.
all = [
    "docling>=2.0", "pypdfium2>=4.0", "python-docx>=1.0", "openpyxl>=3.1",
    "python-pptx>=1.0", "mcp>=1.0", "rapidocr>=3.0", "onnxruntime>=1.16",
    "sentence-transformers>=3.0", "numpy>=1.24",
    "transformers>=4.40", "timm>=1.0", "torch>=2.0", "zstandard>=0.22",
]
# Everything, on a machine that keeps its own accelerated onnxruntime. The only
# difference is that the CPU runtime is not pinned, so upgrading does not put it
# back over `onnxruntime-gpu`.
all-gpu = [
    "docling>=2.0", "pypdfium2>=4.0", "python-docx>=1.0", "openpyxl>=3.1",
    "python-pptx>=1.0", "mcp>=1.0", "rapidocr>=3.0", "zstandard>=0.22",
    "sentence-transformers>=3.0", "numpy>=1.24",
    "transformers>=4.40", "timm>=1.0", "torch>=2.0",
]

[project.scripts]
mdcx = "mdcx.archive:main"
mdcx-convert = "mdcx.cli:main"
mdcx-search = "mdcx.search:main"

[project.urls]
Homepage = "https://github.com/jorgell23-sys/mdcx"
Repository = "https://github.com/jorgell23-sys/mdcx"
Issues = "https://github.com/jorgell23-sys/mdcx/issues"
Changelog = "https://github.com/jorgell23-sys/mdcx/releases"

[tool.hatch.build.targets.wheel]
packages = ["src/mdcx"]

[tool.hatch.build.targets.sdist]
include = ["src/mdcx", "tests", "tools", "README.md", "CHANGELOG.md", "LICENSE", "NOTICE", "CITATION.cff"]