pyproject.toml

[project]
name = "advanced-rag"
version = "0.1.1"
description = "Advanced RAG Bootcamp: reference implementations and shared library"
authors = [ {name = "Vector AI Engineering", email = "ai_engineering@vectorinstitute.ai"}]
readme = "README.md"
repository = "https://github.com/VectorInstitute/advanced-rag"
requires-python = ">=3.12"
dependencies = [
    "aieng-advanced-rag",
    "filelock==4.2.4",
    "idna>=3.18",
    "jupyter-server>=2.20.0",
    "jupyterlab>=4.6.2",
    "mistune>=3.3.3",
    "nbconvert>=7.17.0",
    "pip>=26.2",
    "pygments>=2.20.0",
    "requests>=2.34.2",
    "tornado>=6.5.7",
    "urllib3>=2.7.0",
    "virtualenv>=21.7.13",
]

[dependency-groups]
dev = [
    # The platform's `onboard` CLI. It runs tests/test_integration.py with this
    # venv's interpreter, so the gate can import aieng.advanced_rag.
    "aieng-platform-onboard>=0.7.3",
    "codecov>=2.1.13",
    "mypy>=2.2.0",
    "nbqa>=1.9.1",
    "pip-audit>=2.10.1",
    "pre-commit>=4.6.0",
    "pytest>=9.1.1",
    "pytest-asyncio>=1.4.0",
    "pytest-cov>=6.0.0",
    "pytest-mock>=3.14.0",
    "ruff>=0.15.21",
]

[tool.uv]
# msgpack 1.2.3 and authlib 1.9.0 fix CVE-2026-114613 (a leak and a stale
# reference in the C unpacker) and CVE-2026-96760 (a JWS signature bypass).
constraint-dependencies = [
    "msgpack>=1.2.3",
]
# aieng-platform-onboard pins four of its dependencies exactly, below the
# versions this project already resolves: filelock 3.20.3, virtualenv 20.36.1,
# urllib3 2.7.0, and authlib 1.6.12. The overrides replace those pins so the
# onboarding CLI shares the venv without downgrading anything. Remove them when
# the CLI relaxes its pins. docling-slim 2.133.0 caps pillow below 13, and
# pillow 12.3.0 carries eight advisories fixed in 13.0.0 (PDF append, TIFF
# tile decoding, morphology). Remove the pillow override when docling-slim
# allows 13.
override-dependencies = [
    "authlib>=1.9.0,<2",
    "filelock==4.2.4",
    "pillow>=13.0.0",
    "urllib3>=2.7.0,<3",
    "virtualenv>=21.7.13",
]

[tool.uv.workspace]
members = [
  "aieng-advanced-rag",
]

# PyTorch comes from its CPU wheel index on Linux. The PyPI wheel pulls 15
# nvidia-* packages, 2.7 GB of CUDA libraries that a CPU workspace never loads,
# and nothing in the reference implementations needs a GPU: bge-small-en-v1.5
# runs on CPU. macOS and Windows already get CPU wheels from PyPI. A GPU
# workspace that needs CUDA torch can install it into its own venv.
[[tool.uv.index]]
name = "pytorch-cpu"
url = "https://download.pytorch.org/whl/cpu"
explicit = true

[tool.uv.sources]
aieng-advanced-rag = { workspace = true }
torch = { index = "pytorch-cpu", marker = "sys_platform == 'linux'" }
torchvision = { index = "pytorch-cpu", marker = "sys_platform == 'linux'" }

[tool.mypy]
# Resolve RI2 script imports as the same packages used by the runners.
mypy_path = [
    "$MYPY_CONFIG_FILE_DIR/implementations/cross_document_synthesis",
]
ignore_missing_imports = true
install_types = true
pretty = true
namespace_packages = true
explicit_package_bases = true
non_interactive = true
warn_unused_configs = true
allow_any_generics = false
allow_subclassing_any = false
allow_untyped_calls = false
allow_untyped_defs = false
allow_incomplete_defs = false
check_untyped_defs = true
allow_untyped_decorators = false
warn_redundant_casts = true
warn_unused_ignores = true
warn_return_any = true
implicit_reexport = false
strict_equality = true
extra_checks = true

# ADK is an optional extra, so mypy normally type-checks this repository without
# `google-adk` installed, and `BaseToolset` resolves to `Any`. Subclassing it is
# the documented way to build a toolset, so the check is relaxed for that one
# module rather than silenced with an ignore comment, which would itself become
# an error once the extra is installed.

# docling-core ships py.typed but binds its public names without `__all__`, so
# strict re-export checking rejects `from docling_core.types.doc.document import
# TableItem` for a name the module exposes at runtime. The gap is in the
# package's typing, so it is relaxed for that package alone.
[[tool.mypy.overrides]]
module = "docling_core.*"
implicit_reexport = true

[[tool.mypy.overrides]]
module = ["*.agents.toolset"]
disallow_subclassing_any = false

# RAGAS is an optional RI1 evaluation extra. Its adapter bases resolve to Any
# when that extra is absent, just as ADK's optional toolset base does above.
[[tool.mypy.overrides]]
module = ["advanced_lab.ragas_eval", "*.advanced_lab.ragas_eval"]
disallow_subclassing_any = false

[tool.ruff]
include = ["*.py", "pyproject.toml", "*.ipynb"]
line-length = 120

[tool.ruff.format]
quote-style = "double"
indent-style = "space"
docstring-code-format = true

[tool.ruff.lint]
select = [
    "A", # flake8-builtins
    "B", # flake8-bugbear
    "COM", # flake8-commas
    "C4", # flake8-comprehensions
    "RET", # flake8-return
    "SIM", # flake8-simplify
    "ICN", # flake8-import-conventions
    "Q", # flake8-quotes
    "RSE", # flake8-raise
    "D", # pydocstyle
    "E", # pycodestyle
    "F", # pyflakes
    "I", # isort
    "W", # pycodestyle
    "N", # pep8-naming
    "ERA", # eradicate
    "PL", # pylint
]
fixable = ["A", "B", "COM", "C4", "RET", "SIM", "ICN", "Q", "RSE", "D", "E", "F", "I", "W", "N", "ERA", "PL"]
ignore = [
    "B905", # `zip()` without an explicit `strict=` parameter
    "E501", # line too long
    "D203", # 1 blank line required before class docstring
    "D213", # Multi-line docstring summary should start at the second line
    "PLR2004", # Replace magic number with named constant
    "PLR0913", # Too many arguments
    "COM812", # Missing trailing comma
]

# Ignore import violations in all `__init__.py` files.
[tool.ruff.lint.per-file-ignores]
"__init__.py" = ["E402", "F401", "F403", "F811"]
# Test names under `implementations/` are full sentences, so a per-test
# docstring would restate the name. D100 stays on, so every test module still
# says what it covers. This matches the per-file-ignores the aieng-advanced-rag
# package already sets for its own tests.
"implementations/**/tests/**" = ["D101", "D102", "D103", "D104", "D105"]
# The measurement runners reach for LightRAG, numpy, and the OpenAI client,
# every one of them an optional extra of `aieng-advanced-rag`. Importing them
# at module level would make the runner unimportable without the extras
# installed, which is the same reason the package itself ignores this rule.
# Each measurement directory is listed as it is added.
"implementations/*/graphrag_null/*.py" = ["PLC0415"]
# The onboarding gate imports each optional extra inside the test that needs it,
# so a missing extra fails that one check with a fix instead of failing
# collection. Test names are sentences, as under implementations/.
"tests/**" = ["D101", "D102", "D103", "D104", "D105", "PLC0415"]
# The publishing scripts' tests, named as sentences like the ones above.
"scripts/tests/**" = ["D101", "D102", "D103", "D104", "D105"]

[tool.ruff.lint.pep8-naming]
ignore-names = ["X*", "setUp"]

[tool.ruff.lint.isort]
lines-after-imports = 2

[tool.ruff.lint.pydocstyle]
convention = "numpy"

[tool.ruff.lint.pycodestyle]
max-doc-length = 88

[tool.pytest.ini_options]
markers = [
    "integration_test: marks tests as integration tests",
]