Nomearod/agentbench

1[project]2name = "agent-bench"3version = "0.1.0"4description = "Evaluation-first agentic RAG system built from API primitives"5requires-python = ">=3.11"6dependencies = [7    "anthropic>=0.40.0",8    "openai>=1.50.0",9    "fastapi>=0.115.0",10    "uvicorn[standard]>=0.30.0",11    "pydantic>=2.9.0",12    "pydantic-settings>=2.5.0",13    "python-dotenv>=1.0.0",14    "pyyaml>=6.0",15    "sentence-transformers>=3.0.0,<5.0.0",16    "faiss-cpu>=1.8.0",17    "rank-bm25>=0.2.2",18    "structlog>=24.0.0",19    "httpx>=0.27.0",20    "simpleeval>=1.0.0",21    "numpy>=1.26.0",22    "psutil>=5.9.0",23    "langchain>=0.2.0,<1.0.0",24    "langchain-openai>=0.1.0",25    "langchain-anthropic>=0.1.0",26]27 28[project.optional-dependencies]29dev = [30    "pytest>=8.0.0",31    "pytest-asyncio>=0.24.0",32    "ruff>=0.6.0",33    "mypy>=1.11.0",34    "respx>=0.21.0",35    "types-PyYAML",36    "agent-bench[stats]",37]38stats = [39    "scipy>=1.11.0",40    "pandas>=3.0,<4",41]42plots = [43    # Only needed to regenerate README figures (`make plots`); the CI freshness44    # check is hash-based and matplotlib-free, so this stays out of [dev].45    "matplotlib>=3.8,<4",46]47modal = [48    "modal>=0.66.0",49]50ner = [51    "spacy>=3.7.0",52]53 54[tool.setuptools.packages.find]55include = ["agent_bench*"]56 57[build-system]58requires = ["setuptools>=69.0"]59build-backend = "setuptools.build_meta"60 61[tool.pytest.ini_options]62asyncio_mode = "auto"63testpaths = ["tests"]64 65[tool.ruff]66target-version = "py311"67line-length = 10068extend-exclude = ["scripts/_dev"]69 70[tool.ruff.lint]71select = ["E", "F", "I", "N", "W"]72 73[tool.mypy]74python_version = "3.11"75warn_return_any = true76warn_unused_configs = true77exclude = ["scripts/_dev/"]78