pyproject.toml

[build-system]
requires = ["setuptools>=68"]
build-backend = "setuptools.build_meta"

[project]
name = "lexverse"
version = "1.0"
description = "Task-oriented legal benchmark runtime for legal agents"
license = { text = "Apache-2.0" }
requires-python = ">=3.10"
dependencies = [
    "pydantic>=2.0.0",
    "pyyaml>=6.0",
    "filelock>=3.0",
]

[project.optional-dependencies]
# Model backends.
openai = [
    "openai>=1.0",
]
local-transformers = [
    "openai>=1.0",
    "torch>=2.2",
    # Shared generation API for original Qwen remote code and native Qwen3.
    "transformers==4.51.3",
    "einops>=0.8,<0.9",
    "transformers-stream-generator==0.0.5",
    "tiktoken>=0.5.2",
    "sentencepiece==0.2.0",
    "protobuf",
    "accelerate",
    "huggingface_hub>=0.25",
]
# vLLM 0.11.x keeps Transformers 4.x compatibility with the LawBench/LTP stack.
# Install on hardware supported by vLLM; the default Mac smoke uses Transformers/MPS.
local-vllm = [
    "openai>=1.0",
    "vllm>=0.11,<0.12",
    "huggingface_hub>=0.25",
]

# Benchmarks.
dlawbench = [
    "openai>=1.0",
]
j1bench = [
    "openai>=1.0",
    "huggingface_hub>=0.25",
    "jsonlines>=3.0",
    "requests>=2.0",
    "tqdm>=4.0",
    "numpy>=1.20,<2",
]
# rouge and rouge_chinese share an import path; combined installs may change Rouge scores.
lawbench = [
    "rouge_chinese>=1.0",
    "cn2an>=0.5",
    "jieba>=0.42",
    "numpy>=1.20,<2",
    "nltk>=3.6",
    "pandas>=1.3",
    "opencc-python-reimplemented>=0.1.7",
    "timeout-decorator>=0.5",
    "tqdm>=4.0",
    "pypinyin>=0.49",
    "python-Levenshtein>=0.20",
    "ltp>=4.2",
    "torch>=2.0",
    # The supported LTP stack requires Transformers 4.x.
    "transformers>=4.0,<5",
]
legalworld = [
    "huggingface_hub>=0.25",
    "camel-ai==0.2.90",
    # CAMEL 0.2.90 imports FastMCP from mcp.server (removed in MCP 2.x).
    "mcp>=1.3.0,<2",
    "python-dotenv>=1.0.1",
    "docxtpl>=0.16.7",
    "python-docx>=1.1.0",
    "reportlab>=4.2.0",
    "numpy>=1.26",
]
lexeval = [
    "jsonlines>=3.0",
    "jieba>=0.42",
    "rouge>=1.0",
    # Upstream evaluators require NumPy 1.x compatibility.
    "numpy>=1.20,<2",
]
plawbench = [
    "openai>=1.0",
    "json-repair>=0.30",
]

# SimuCourt, MASER and LeCaRDv2 benchmark dependencies.
simucourt = [
    "openai>=1.0",
]
maser = [
    "openai>=1.0",
    "numpy>=1.24",
    "tqdm>=4.66",
    "jsonlines>=4.0",
]
lecardv2 = [
    "openai>=1.0",
    "jieba>=0.42",
    "numpy>=1.20",
    "pytrec-eval-terrier>=0.5.6",
]

# Development and testing.
test = [
    "pytest>=7.0",
    "pytest-asyncio>=0.23",
]

[project.scripts]
lexverse = "lexverse.cli:main"

[tool.setuptools.packages.find]
where = ["."]
include = ["lexverse", "lexverse.*"]
exclude = ["reference_harness*", "tests*", "data*", "docs*"]

[tool.setuptools.package-data]
"lexverse.benchmarks.dlawbench" = ["catalog.yaml"]
"lexverse.benchmarks.j1bench" = ["patches/*.patch", "catalog.yaml"]
"lexverse.benchmarks.lawbench" = ["patches/*.patch", "catalog.yaml"]
"lexverse.benchmarks.legalworld" = ["catalog.yaml", "patches/*.patch"]
"lexverse.benchmarks.lexeval" = ["patches/*.patch", "catalog.yaml"]
"lexverse.benchmarks.plawbench" = ["patches/*.patch", "catalog.yaml"]
"lexverse.benchmarks.simucourt" = ["catalog.yaml", "patches/*.patch"]
"lexverse.benchmarks.maser" = ["catalog.yaml", "patches/*.patch"]
"lexverse.benchmarks.lecardv2" = ["catalog.yaml"]

[tool.pytest.ini_options]
testpaths = ["tests"]
asyncio_mode = "auto"
markers = [
    "integration: needs upstream cache and/or network (offline-runnable once cached)",
    "live_api: calls a real paid model API; excluded by default",
    "local_hardware: real local backend requiring suitable hardware and cached weights",
]
# Exclude live API tests by default.
addopts = "-m 'not live_api'"