[project]
name = "lilbee"
version = "0.6.90b442"
description = "The whole local AI stack in one executable: run and manage local AI models across every GPU you have, and search your files, code, and the web pages you crawl, with answers that cite the source. Per-project libraries, semantic and hybrid search, vision OCR, an auto-built wiki, CLI, TUI, MCP server, REST API, and Python library. No model server, no database server, no containers. Works with your own Ollama or LM Studio models too."
license = "MIT"
authors = [{ name = "tobocop2", email = "5562156+tobocop2@users.noreply.github.com" }]
requires-python = ">=3.11"
dynamic = ["readme"]
keywords = [
    "rag", "local-rag", "local-llm", "local-ai",
    "mcp", "mcp-server", "model-context-protocol",
    "semantic-search", "vector-search", "retrieval",
    "embeddings", "knowledge-base", "knowledge-management",
    "huggingface", "ocr", "web-crawler",
    "model-manager", "llama-cpp", "gguf", "multi-gpu", "ollama", "lm-studio",
    "obsidian", "second-brain", "personal-ai", "self-hosted", "privacy", "cli", "tui",
]
classifiers = [
    "Development Status :: 4 - Beta",
    "Environment :: Console",
    "Environment :: Web Environment",
    "Intended Audience :: Developers",
    "Intended Audience :: End Users/Desktop",
    "Intended Audience :: Science/Research",
    "Operating System :: OS Independent",
    "Programming Language :: Python :: 3",
    "Programming Language :: Python :: 3.11",
    "Programming Language :: Python :: 3.12",
    "Programming Language :: Python :: 3.13",
    "Topic :: Scientific/Engineering :: Artificial Intelligence",
    "Topic :: Text Processing :: Indexing",
    "Topic :: Software Development :: Libraries :: Python Modules",
    "Typing :: Typed",
]

dependencies = [
    # pylance has no macOS x86_64 wheel; requiring the extra there backtracks
    # lancedb to 0.25.3, which predates list_tables(). Intel Macs run plain
    # lancedb (the +compat wheel); import lance is confined to the multi-GPU
    # shard-adoption path, which that hardware cannot reach.
    "lancedb[pylance] ; sys_platform != 'darwin' or platform_machine != 'x86_64'",
    "lancedb ; sys_platform == 'darwin' and platform_machine == 'x86_64'",
    "xberg>=1.1.2",
    # The engine user-lock protocol leans on release(), is_singleton reentrancy
    # and the SoftFileLock fallback; the ceiling is the line verified against it.
    "filelock<3.33",
    "tree-sitter-language-pack>=1.8.0,<2.0",
    "typer>=0.12",
    "tiktoken",
    "textual>=0.75",
    # Floor is 2.0: the tool server needs mcp.server.mcpserver, which 1.x lacks.
    "mcp>=2.0.0,<3",
    "pillow>=11.3.0",
    "litestar>=2.0",
    "uvicorn>=0.30",
    "httpx",
    "diskcache>=5.6.1",
    "jinja2>=2.11.3",
    "typing-extensions>=4.5.0",
    "huggingface_hub>=1.27.0",
    "psutil>=5.9",
    "pydantic-settings>=2.14.2",
    # 2.5 drops the 3.11 floor and ships 3.12-only stubs mypy rejects at the 3.11 target.
    "numpy<2.5",
    # Entity schema patterns are LLM-authored, so their matches need a wall-clock
    # bound; regex takes a per-match timeout where the stdlib re module cannot.
    # Already present transitively via tiktoken.
    "regex",
    # The read side is stdlib tomllib; this is its write half.
    "tomli-w>=1.0",
    "cachetools>=7.1.4",
    # gitignore-syntax matching for .lilbeeignore. Already present transitively;
    # the floor is where the non-deprecated "gitignore" pattern factory lands.
    "pathspec>=1.1.1",
]

[project.optional-dependencies]
# The bundled llama-server and gguf-parser. Not a hard dependency because the
# distribution is published on lilbee.sh rather than PyPI, and a hard reference
# to a name PyPI answers 404 for made the documented install command backtrack
# to a months-old release that predated it. Installing without this leaves a
# working lilbee that says what to install the moment it needs an engine.
# The floor is below every published engine wheel (they carry lilbee's own
# version), so it never excludes a real release; its jobs are to keep the
# in-repo 0.0.0 stub from satisfying an index install and, because the
# specifier names a pre-release, to let plain `pip` resolve the beta-versioned
# wheels without `--pre`.
engine = ["lilbee-engine>=0.6.90b1"]
litellm = ["litellm>=1.84.0,<1.99"]
graph = ["spacy>=3.8", "graspologic-native>=1.2"]
crawler = ["crawl4ai>=0.9.0"]
# A fallback source of the CUDA 12 runtime, from NVIDIA's official PyPI wheels;
# fleet.cuda_runtime adds their lib dirs to the engine's LD_LIBRARY_PATH. The engine
# wheel now ships cudart/cublas/cublasLt beside llama-server on both Linux and Windows
# (build_llama_server.sh), so this is only needed for an engine built elsewhere.
# Linux x86_64 only: macOS and aarch64 have no CUDA build to feed.
cuda12 = [
    "nvidia-cuda-runtime-cu12; platform_system=='Linux' and platform_machine=='x86_64'",
    "nvidia-cublas-cu12; platform_system=='Linux' and platform_machine=='x86_64'",
    "nvidia-cuda-nvrtc-cu12; platform_system=='Linux' and platform_machine=='x86_64'",
]
release = ["lilbee[crawler,litellm,graph]"]

[project.urls]
Homepage = "https://lilbee.sh/"
Repository = "https://github.com/tobocop2/lilbee"
Issues = "https://github.com/tobocop2/lilbee/issues"

[project.scripts]
lilbee = "lilbee.runtime.launcher:main"

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

# The README shows each demo as a GitHub video player, which only GitHub
# renders. hatch_build.py rewrites those back to the gh-pages GIFs so the PyPI
# page keeps its moving pictures instead of a wall of links.
[tool.hatch.metadata.hooks.custom]

[tool.hatch.build.targets.wheel]
packages = ["src/lilbee"]

# Dev resolves the bundled engine from the in-repo wheel package; CI fills its
# bin/ with the self-contained llama-server + llama-swap + gguf-parser before
# building. Release installs (`pip install lilbee`) resolve `lilbee-engine` from
# the index instead -- this source override is workspace metadata, not published deps.
[tool.uv.sources]
lilbee-engine = { path = "packaging/engine-wheel" }

# `uv pip install <checkout>[engine] --extra-index-url ...` applies the source
# above no matter the specifier (uv drops version bounds on sourced packages),
# so it silently installed the binariless stub where the deployment wanted the
# real engine wheel. This turns sources off for the `uv pip` interface only:
# that command now resolves lilbee-engine from the index, while `uv sync` /
# `uv lock` (the project interface) keep using the stub.
[tool.uv.pip]
no-sources = true

[dependency-groups]
dev = [
    # The bundled engine is an optional extra for users, because the wheel is
    # published on lilbee.sh rather than PyPI and a hard reference to a name
    # PyPI answers 404 for breaks the documented install. Development and CI
    # still need it: the pin test reads its metadata, and the local source
    # below makes it a workspace path rather than an index lookup.
    "lilbee-engine",
    # The self-check-extras charset probe and its tests call chardet directly;
    # without this, chardet only arrives via the crawler extra, and the bare
    # `uv sync --locked` CI test env cannot run those tests.
    "chardet>=7",
    # llama.cpp's Python package, for tests only. It is not a second GGUF reader
    # in the product: gguf-parser is, built from the pin that builds
    # llama-server. This supplies GGML_QUANT_SIZES so the measured
    # bytes-per-weight table is checked against the authoritative block sizes.
    "gguf>=0.18",
    "hypothesis>=6.100",
    "mypy>=1.19.1",
    "packaging>=24",
    "pytest",
    "pytest-asyncio",
    "pytest-cov",
    "reportlab",
    "ruff>=0.15.4",
    "httpx>=0.27",
    "pytest-httpserver",
    "types-pyyaml>=6.0.12.20250915",
    "pytest-xdist>=3.8.0",
    "pytest-forked>=1.6.0",
    "pytest-timeout>=2.3",
    "pre-commit>=4.0",
    "types-cachetools>=7.0.0.20260713",
]

[tool.pytest.ini_options]
pythonpath = [".", "tests"]
testpaths = ["tests"]
asyncio_mode = "auto"
addopts = "--ignore=tests/integration"
timeout = 60
markers = [
    "slow: integration tests that download models or take >10s",
    "real_model_classify: opt out of the global _classify_installed_models mock",
    "real_litellm_probe: opt out of the global litellm_available True override",
    "real_engine_resolution: opt out of the engine-binary seal (host PATH and LILBEE_LLAMA_SERVER_PATH resolve again)",
    "xdist_group: pin grouped tests to a single xdist worker for serial execution",
    "no_health_default: opt out of the launch tests' default-healthy fixture",
    "no_warm_default: opt out of the launch tests' default chat-warm fixture",
    "live_picks: opt out of the seeded SAMPLE_PICKS and resolve picks for real",
    "real_hf_client: opt out of the global hf_client.fetch_models stub (tests the client itself)",
]

[tool.coverage.run]
source = ["lilbee"]
concurrency = ["thread"]
# Generated modules are machine output, not code anyone wrote or can meaningfully
# test; measuring them says nothing about the suite. Excluded so the number stays
# a statement about hand-written code, and so regenerating cannot move it.
omit = ["src/lilbee/_generated/*"]

[tool.coverage.report]
show_missing = true
fail_under = 100
exclude_lines = [
    "pragma: no cover",
    "if TYPE_CHECKING:",
    "class.*Protocol",
]

[tool.ruff]
target-version = "py311"
line-length = 100
src = ["src", "tests", "tools"]

[tool.ruff.lint]
select = [
    "E",     # pycodestyle errors
    "W",     # pycodestyle warnings
    "F",     # pyflakes
    "I",     # isort
    "UP",    # pyupgrade
    "B",     # flake8-bugbear
    "SIM",   # flake8-simplify
    "RUF",   # ruff-specific
    "S",     # flake8-bandit: security antipatterns (eval, exec, pickle, unsafe subprocess, etc.)
    "C90",   # mccabe complexity
    "N",     # pep8-naming
    "RET",   # flake8-return: consistent return semantics
    "TID",   # flake8-tidy-imports: forbid relative imports
    "PLR0911", # too many returns
    "PLR0912", # too many branches
    "PLR0915", # too many statements
    "PLR2004", # magic value used in comparison
]
# DeepSeek V3.1's tool-call markers are intentionally fullwidth (｜ U+FF5C,
# ▁ U+2581). These are the exact characters the model emits; substituting
# the ASCII look-alikes would break extraction.
allowed-confusables = ["｜", "▁"]

[tool.ruff.lint.mccabe]
max-complexity = 10

[tool.ruff.lint.flake8-bugbear]
# litestar's Body(), like Parameter() (already in ruff's B008 default allowlist),
# is a declarative parameter marker that is safe as an argument default.
extend-immutable-calls = ["litestar.params.Body"]

[tool.ruff.lint.per-file-ignores]
# Test-context S-rule noise. Each of these is either a test fixture
# (hardcoded token/password literals, fake /tmp paths, 0.0.0.0 bind
# assertions), or an intentional subprocess spawn for an e2e test:
# - S101 assert (tests are built on it)
# - S104 bind-all-interfaces (testing host parsing)
# - S105/S106 hardcoded password string/arg (test fixtures)
# - S108 /tmp paths (mocked, not actual temp-file usage)
# - S311 pseudo-random (seeded generators building test data, never a secret;
#   the seed is the point, since a load test that cannot be replayed is useless)
# - S603/S607 subprocess (e2e tests spawning sys.executable)
# Test files also ignore PLR2004 (test fixtures use literal expected
# values inline; extracting them to constants hurts readability) and
# the pep8-naming rules (test helpers frequently shadow PascalCase
# class names from the system under test).
"tests/**/*.py" = [
    "S101", "S104", "S105", "S106", "S108", "S311", "S603", "S607",
    "PLR2004",
    "N802", "N806", "N818", "N817",
]
# QA matrix runner is a test framework just like tests/; same ignores apply.
"tools/qa/**/*.py" = [
    "S101", "S104", "S105", "S106", "S108", "S603", "S607",
    "PLR2004",
    "N802", "N806", "N818", "N817",
]
# Surface-QA harnesses drive the CLI/TUI/HTTP surfaces like tests do; same ignores
# (S311 for the same reason as tests: seeded fuzz walks must replay).
"scripts/qa/**/*.py" = [
    "S101", "S104", "S105", "S106", "S108", "S311", "S603", "S607",
    "PLR2004",
    "N802", "N806", "N818", "N817",
]
# Invokes the trusted `gh` CLI with controlled args to build the compat index.
"tools/build_compat_index.py" = ["S603", "S607"]
"conftest.py" = ["S101"]

[tool.mypy]
python_version = "3.11"
warn_return_any = true
warn_unused_configs = true
disallow_untyped_defs = true
check_untyped_defs = true
ignore_missing_imports = true