pyproject.toml

# Project metadata + tool config.
#
# Three-layer dependency design (don't drift these):
#   - ``[project.dependencies]`` below is the **authoritative** runtime
#     dep set. ``pip install .`` (or ``pip install -e .``) installs the
#     package and these deps in one shot.
#   - ``requirements.txt`` is a dev-convenience mirror (loose lower
#     bounds matching this file) plus dev tooling (``pytest``).
#   - ``requirements-lock.txt`` is a pip-freeze-style pinned lockfile
#     for the current security-refreshed CPU dev environment. Historical
#     report numbers remain anchored to the frozen release tag documented
#     in ``requirements-lock.txt``.
#
# To get a working dev environment use ``make install`` — that runs both
# ``pip install -r requirements.txt`` (for ``pytest`` etc.) and
# ``pip install -e .``.

[build-system]
requires = ["setuptools>=61", "wheel"]
build-backend = "setuptools.build_meta"

[project]
name = "msmarco-genqa"
version = "0.1.0"
description = "Retrieval-augmented question answering on MS MARCO."
readme = "README.md"
requires-python = ">=3.10"
license = { file = "LICENSE" }
authors = [{ name = "Gioia Zheng", email = "gioia.zheng.stud@gmail.com" }]
keywords = ["rag", "information-retrieval", "msmarco", "reranking", "evaluation"]

# Runtime deps. Loose lower bounds mirroring ``requirements.txt``; exact
# pins live in ``requirements-lock.txt``. Edit all three together.
dependencies = [
    # Core scientific stack
    "numpy>=1.24",
    "pandas>=2.0",
    "matplotlib>=3.7",
    # Configuration
    "pyyaml>=6.0",
    # Datasets and IO
    "datasets>=2.14",
    "ir-datasets>=0.5.5",
    "lxml>=6.1.0",
    # Lexical retrieval
    "bm25s>=0.2",
    # Dense retrieval
    "sentence-transformers>=5.5.1",
    "faiss-cpu>=1.7",
    # Generation
    "torch>=2.12.0",
    "transformers>=5.10.2",
    "sentencepiece>=0.1.99",
    # Evaluation
    "evaluate>=0.4",
    "rouge_score>=0.1.2",
    "sacrebleu>=2.3",
    "bert_score>=0.3.13",
]

[project.urls]
Repository = "https://github.com/GioiaZheng/msmarco-genqa"
Results = "https://github.com/GioiaZheng/msmarco-genqa/blob/main/RESULTS.md"
Reproducibility = "https://github.com/GioiaZheng/msmarco-genqa/blob/main/REPRODUCIBILITY.md"

[project.optional-dependencies]
serve = [
    "fastapi>=0.110",
    "uvicorn>=0.27",
]
tracking = [
    "mlflow>=2.10",
    "wandb>=0.16",
]
evaluation = [
    "ir-measures>=0.3.3",
]

[tool.setuptools.packages.find]
# Two roots: ``src/`` for the library package and the repo root for the
# ``experiments`` entry-point layer. The CLI shims under
# ``msmarco_genqa.cli`` re-export ``experiments.run_*.main`` so the
# ``mgq-*`` console scripts can find them; that requires ``experiments``
# itself to be installed.
where = ["src", "."]
include = ["msmarco_genqa*", "experiments"]
exclude = ["tests*", "scripts*"]

[project.scripts]
# Console-script entry points. Each shim under msmarco_genqa.cli.<name>
# re-exports main() from the matching experiments/run_<name>.py.
mgq-retrieve = "msmarco_genqa.cli.retrieve:main"
mgq-dense    = "msmarco_genqa.cli.dense:main"
mgq-fuse     = "msmarco_genqa.cli.fuse:main"
mgq-transform-queries = "msmarco_genqa.cli.query_transform:main"
mgq-query-transform-ablation = "msmarco_genqa.cli.query_transform_ablation:main"
mgq-retrieval-report = "msmarco_genqa.cli.retrieval_report:main"
mgq-context-packing-report = "msmarco_genqa.cli.context_packing_report:main"
mgq-rag-triad = "msmarco_genqa.cli.rag_triad:main"
mgq-sweep-summary = "msmarco_genqa.cli.sweep_summary:main"
mgq-export-rag-observatory = "msmarco_genqa.cli.rag_observatory_export:main"
mgq-export-rag-observatory-sweep = "msmarco_genqa.cli.rag_observatory_sweep_export:main"
mgq-rerank   = "msmarco_genqa.cli.rerank:main"
mgq-generate = "msmarco_genqa.cli.generate:main"
mgq-pipeline = "msmarco_genqa.cli.pipeline:main"
mgq-serve    = "msmarco_genqa.service.app:main"
rag-eval     = "msmarco_genqa.cli.rag_eval:main"
mgq-trec-eval = "msmarco_genqa.cli.trec_eval:main"
mgq-trec-release = "msmarco_genqa.cli.trec_release:main"
mgq-beir-release = "msmarco_genqa.cli.beir_release:main"
mgq-nfcorpus-video-release = "msmarco_genqa.cli.nfcorpus_video_release:main"


# --------------------------------------------------------------------------- #
# pytest
# --------------------------------------------------------------------------- #

[tool.pytest.ini_options]
minversion = "7.0"
testpaths = ["tests"]

# Default: only fast unit tests. ``slow`` and ``integration`` are opt-in via
# ``-m slow`` / ``-m integration`` (or via ``make test-slow``).
addopts = "-m 'not slow and not integration'"

markers = [
    "slow: tests that need real models, large data, or network access (HF Hub, ir_datasets). Excluded by default.",
    "integration: end-to-end tests that hit real external services. Excluded by default.",
]


# --------------------------------------------------------------------------- #
# ruff
# --------------------------------------------------------------------------- #
#
# Conservative ruleset: we want catch-real-bugs (pyflakes ``F``) plus
# import-order hygiene (``I``), with a couple of widely-ignored style rules.
# We deliberately do NOT enable opinionated style families (UP, B, S, …) on
# this first pass — they would touch dozens of files without changing
# behaviour and obscure the real fixes.

[tool.ruff]
line-length = 100
target-version = "py310"
extend-exclude = [
    "data",
    "outputs",
    "reports/generated",
    "figures",
]

[tool.ruff.lint]
# Conservative first pass: pyflakes (real bugs only) + a handful of
# whitespace warnings. Import-order (``I``) is intentionally OFF — it would
# rewrite import blocks across many files for zero behavioural change, and
# the user has asked us not to introduce wholesale formatting churn.
# Promote to ``E`` / ``I`` / ``UP`` in a follow-up pass once the codebase
# has been intentionally cleaned.
select = ["F", "W"]
ignore = [
    "E501",  # line-too-long: requirements.txt comments + long URLs in docstrings
    "E731",  # lambda assignment: occasionally clearer than def for one-liners
]