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

[project]
name = "webfetch-llm"
version = "0.1.3"
description = "Own your LLM's web search: a local search -> fetch -> rank pipeline that replaces hosted web_search tools at a fraction of the cost"
readme = "README.md"
requires-python = ">=3.10"
license = { text = "MIT" }
authors = [{ name = "Rishi Gulati" }]
keywords = [
    "llm", "web-search", "agents", "tool-use", "rag",
    "anthropic", "openai", "mcp", "semantic-cache",
]
classifiers = [
    "Development Status :: 4 - Beta",
    "Intended Audience :: Developers",
    "License :: OSI Approved :: MIT License",
    "Programming Language :: Python :: 3.10",
    "Programming Language :: Python :: 3.11",
    "Programming Language :: Python :: 3.12",
    "Topic :: Internet :: WWW/HTTP :: Indexing/Search",
    "Topic :: Scientific/Engineering :: Artificial Intelligence",
]

# Core deps kept lean - no heavy ML libs by default. Cache uses stdlib
# sqlite3. Ranking degrades to BM25 (rank-bm25, pure Python) without the
# rerank extra. readability-lxml and newspaper4k are core because they run
# in the DEFAULT extraction chain - the benchmark numbers assume them.
dependencies = [
    "trafilatura>=1.8",
    "readability-lxml>=0.8",
    "newspaper4k>=0.9",
    "rank-bm25>=0.2.2",
    "requests>=2.31",
    # ddgs is the maintained successor of duckduckgo-search; ddg.py imports it.
    "ddgs>=6.0",
    "anthropic>=0.25",
    "mcp>=1.0",
    # Entry points (webfetch-mcp/-status/-savings) load .env from cwd;
    # the library itself never reads it. See webfetch/_env.py.
    "python-dotenv>=1.0",
]

[project.optional-dependencies]
# Semantic reranking, semantic query cache, compression scoring
rerank = [
    "sentence-transformers>=3.0",
    "numpy>=1.26",
]

# JS-rendered page fallback (playwright must also run `playwright install`)
browser = [
    "playwright>=1.44",
]

# PDF datasheet extraction
pdf = [
    "pdfplumber>=0.11",
]

# HTML table extraction (pandas.read_html appended as markdown;
# tabulate is what DataFrame.to_markdown actually calls)
tables = [
    "pandas>=2.0",
    "tabulate>=0.9",
]

# Kept as a no-op alias (mcp moved to core in 0.1.1 so uvx
# one-liners work); harmless for existing install commands.
mcp = []

# Full install
all = ["webfetch-llm[rerank,browser,pdf,tables,mcp]"]

# Development: tests, lint, .env loading for examples/evals
dev = [
    "pytest>=8.0",
    "ruff>=0.4",
]

[project.urls]
Homepage = "https://github.com/firish/webfetch"
Repository = "https://github.com/firish/webfetch"
Issues = "https://github.com/firish/webfetch/issues"

[project.scripts]
webfetch-mcp = "webfetch.mcp:main"
webfetch-savings = "webfetch.receipts:main"
webfetch-status = "webfetch.status:main"
# Alias named after the PyPI package so `uvx webfetch-llm` starts the MCP
# server directly - uvx runs the script matching the package name, and
# MCP registry clients build run commands from the package identifier.
webfetch-llm = "webfetch.mcp:main"

[tool.hatch.build.targets.wheel]
packages = ["webfetch"]

[tool.ruff]
line-length = 88
target-version = "py310"

[tool.ruff.lint]
# E/F: pycodestyle+pyflakes defaults; I: import sorting; UP: modernize.
select = ["E", "F", "I", "UP"]
# E501 in evals: report tables and URLs run long; the library itself complies.
[tool.ruff.lint.per-file-ignores]
"evals/*" = ["E501"]
"examples/*" = ["E501"]

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