[project]
name = "mempalace"
version = "3.12.0"
description = "Give your AI a memory — mine projects and conversations into a searchable palace. No API key required."
readme = "README.md"
requires-python = ">=3.10"
license = "MIT"
authors = [
    {name = "milla-jovovich"},
]
keywords = [
    "ai", "memory", "llm", "rag", "chromadb", "mcp",
    "vector-database", "claude", "chatgpt", "embeddings",
]
classifiers = [
    "Development Status :: 4 - Beta",
    "Environment :: Console",
    "Intended Audience :: Developers",
    "Programming Language :: Python :: 3",
    "Programming Language :: Python :: 3.10",
    "Programming Language :: Python :: 3.11",
    "Programming Language :: Python :: 3.12",
    "Programming Language :: Python :: 3.13",
    "Programming Language :: Python :: 3.14",
    "Topic :: Scientific/Engineering :: Artificial Intelligence",
    "Topic :: Utilities",
]
dependencies = [
    "chromadb>=1.5.4,<2",
    "pyyaml>=6.0,<7",
    "tomli>=2.0.0; python_version < '3.11'",
    # Multilingual embedding model deps (embeddinggemma-300m ONNX), kept as
    # core deps rather than optional extras so choosing the multilingual model
    # needs no extra pip command. The 300 MB ONNX model itself is
    # lazy-downloaded on first use, not at install time.
    "huggingface_hub>=0.20",
    "tokenizers>=0.15",
    "numpy>=1.24",
    # Tier 6a content-date extraction uses dateutil for natural-language
    # date parsing (filename + content body branches). It happens to install
    # transitively today via chromadb→kubernetes→python-dateutil, but that
    # chain is not a contract — declared explicitly here to ensure
    # _extract_content_date keeps working if upstream kubernetes drops the
    # dependency. Per PR #1584 review (Igor, 2026-05-22).
    "python-dateutil>=2.8",
]

[project.urls]
Homepage = "https://github.com/MemPalace/mempalace"
Repository = "https://github.com/MemPalace/mempalace"
"Bug Tracker" = "https://github.com/MemPalace/mempalace/issues"

[project.scripts]
mempalace = "mempalace.cli:main"
mempalace-mcp = "mempalace.mcp_proxy:main"
mempalace-light-mcp = "mempalace.mcp_light_server:main"

[project.entry-points."mempalace.backends"]
chroma = "mempalace.backends.chroma:ChromaBackend"
milvus = "mempalace.backends.milvus:MilvusBackend"
pgvector = "mempalace.backends.pgvector:PgVectorBackend"
qdrant = "mempalace.backends.qdrant:QdrantBackend"
sqlite_exact = "mempalace.backends.sqlite_exact:SQLiteExactBackend"
rust_exact = "mempalace.backends.rust_exact:RustExactBackend"

# RFC 002 source-adapter entry-point group. Core publishes no first-party
# adapters under this group yet; ``miner.py`` and ``convo_miner.py`` migrate
# onto ``BaseSourceAdapter`` in a follow-up PR. Third-party adapter packages
# (``mempalace-source-cursor``, ``mempalace-source-git``, …) register here.
[project.entry-points."mempalace.sources"]

[project.optional-dependencies]
dev = [
    "pytest>=7.0",
    "pytest-cov>=4.0",
    # Offline media tests exercise real image objects and audio resampling
    # without installing the optional model or its PyTorch dependencies.
    "pillow>=10.0",
    "scipy>=1.10",
    # Retries known-transient ChromaDB-on-Windows HNSW compaction failures in CI
    # (wired only into the test-windows job in ci.yml, scoped to the specific
    # error). Local/Linux/macOS runs never rerun, so real failures stay loud.
    "pytest-rerunfailures>=12.0",
    "ruff==0.16.6",
    "psutil>=5.9",
    # Property-based testing — generates hundreds of random inputs per
    # test to find counterexamples the hand-written positive tests miss.
    # Used opt-in (sprinkle ``@given(...)`` on a test); zero runtime
    # cost on tests that don't use it.
    "hypothesis>=6.0",
    # Pre-commit framework — gates every local ``git commit`` on the
    # same ruff checks CI runs. Required activation step in
    # CONTRIBUTING.md: ``pre-commit install`` once per clone. Without
    # this, contributors silently bypass the lint gate locally and only
    # discover failures at CI time.
    "pre-commit>=3.0",
    # mypy — static type checker. Catches lazy ``list`` / ``dict``
    # annotations that should be ``list[str]`` / ``dict[str, int]``,
    # missing return types, mismatched signatures across call sites.
    # Run ad-hoc with ``mypy mempalace/``; CI doesn't gate on it yet but
    # contributors should fix mypy errors they see before pushing.
    "mypy>=1.0",
]
spellcheck = ["autocorrect>=2.0"]
# Opt-in Milvus backend. Defaults to per-palace Milvus Lite
# (`<palace>/milvus.db`); set MEMPALACE_MILVUS_URI and
# MEMPALACE_MILVUS_TOKEN to point at Milvus server or Zilliz Cloud.
milvus = [
    "pymilvus>=3.0.0",
    "milvus-lite>=3.0; sys_platform != 'win32'",
]
# Opt-in Postgres + pgvector backend. Only the psycopg driver is needed on the
# client; the server must have the `vector` extension available. Selected via
# MEMPALACE_BACKEND=pgvector / --backend pgvector; never required for the
# default (Chroma) install.
pgvector = ["psycopg[binary]>=3.1"]
# Hardware acceleration for the ONNX embedding model. Install exactly one:
#   pip install mempalace[gpu]       — NVIDIA CUDA
#   pip install mempalace[dml]       — DirectML (Windows AMD/Intel/NVIDIA)
#   pip install mempalace[coreml]    — macOS Neural Engine
# After install, set MEMPALACE_EMBEDDING_DEVICE=cuda|dml|coreml (or "auto").
gpu = ["onnxruntime-gpu>=1.16"]
dml = ["onnxruntime-directml>=1.16"]
coreml = ["onnxruntime>=1.16"]
# Opt-in PyTorch/Transformers implementation of Google EmbeddingGemma 2.
# The upstream image extra supplies Pillow and torchvision, which the
# processor imports even for text-only use.
embeddinggemma2 = [
    "sentence-transformers[image]>=6.1.0",
    "transformers>=5.19.0",
    "torch>=2.14.0",
    # Pillow < 12.3.0 has heap out-of-bounds writes and other memory-safety
    # advisories reachable from untrusted images.
    "pillow>=12.3.0",
]
multimodal = [
    "sentence-transformers[image]>=6.1.0",
    "transformers>=5.19.0",
    "torch>=2.14.0",
    # Pillow < 12.3.0 has heap out-of-bounds writes and other memory-safety
    # advisories reachable from untrusted images.
    "pillow>=12.3.0",
    "soundfile>=0.13.1",
    "scipy>=1.10",
    "torchcodec>=0.17.0",
]
# Multilingual extra kept as a no-op alias for back-compat with
# `pip install mempalace[multilingual]` — these deps now ship in core.
multilingual = []
# Binary-format extraction for `mempalace mine --mode extract`. Per-format
# transformer routing: striprtf for .rtf, MarkItDown for .pdf/.docx/.pptx/
# .xlsx/.epub.
extract = [
    "striprtf>=0.0.27",
    # Per PR #1555 review (Igor): bare ``markitdown`` is enough to import
    # but not enough to convert — actual PDF/DOCX/PPTX/XLSX extraction
    # needs MarkItDown's per-format sub-extras, which pull pdfminer-six,
    # mammoth, python-pptx, openpyxl, etc. EPUB is handled by base
    # markitdown (EpubConverter uses beautifulsoup4, a base requirement),
    # so no [epub] sub-extra is needed (and none exists in 0.1.5). RTF is
    # covered by striprtf above.
    "markitdown[docx,pdf,pptx,xlsx]>=0.1.5",
]

[dependency-groups]
dev = [
    "pytest>=7.0",
    "pytest-cov>=4.0",
    "pillow>=10.0",
    "scipy>=1.10",
    # Retries known-transient ChromaDB-on-Windows HNSW compaction failures in CI
    # (wired only into the test-windows job in ci.yml, scoped to the specific
    # error). Local/Linux/macOS runs never rerun, so real failures stay loud.
    "pytest-rerunfailures>=12.0",
    "ruff==0.16.6",
    "psutil>=5.9",
    "hypothesis>=6.0",
    "pre-commit>=3.0",
    "mypy>=1.0",
]

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

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

[tool.ruff]
line-length = 100
target-version = "py310"
# "*.md": ruff 0.16 began formatting Python inside markdown fences, which
# pulls ~96 docs into scope and reflows hand-aligned example code (the
# column-aligned field comments in docs/rfcs/002 are deliberate). Docs are
# authored prose, not linted source — keep the formatter on real modules.
extend-exclude = ["benchmarks", "*.md"]

[tool.ruff.lint]
select = ["E", "F", "W", "C901", "I", "RUF100", "B", "RET", "ERA"]
# RET501: an explicit `return None` in a no-op default or a test fake shows
# the reader it stands in for a call that returns a value.
# RET503: every hit ends in a call that always raises, a retry loop whose last
# attempt re-raises, or an endless generator; an added `return` would be dead.
# B905: fires on every zip() once the target is py310. Choosing strict= per
# call is its own lint stage, not part of raising the Python floor.
ignore = ["E501", "RET501", "RET503", "B905"]
# Codes of rules planned for later lint stages. Their noqa marks are
# deliberate, so RUF100 must not strip them before those rules are on.
external = ["A002", "ARG001", "ARG002", "BLE001", "S310"]

[tool.ruff.lint.isort]
known-first-party = ["mempalace"]
# Keep `x as y` names in the statement they came from, so a noqa on that
# statement still covers them.
combine-as-imports = true

[tool.ruff.lint.per-file-ignores]
# mcp_server fragments exec into the package namespace. Names they use live
# in __init__.py and earlier fragments, so F821 is expected; __init__.py
# imports those names for the fragments, so F401 is expected there too.
"mempalace/mcp_server/__init__.py" = ["F401"]
"mempalace/mcp_server/_*.py" = ["F821"]
"mempalace/mcp_server/tools_*.py" = ["F821"]
"mempalace/mcp_server/schemas.py" = ["F821"]
"mempalace/mcp_server/protocol.py" = ["F821"]
"mempalace/mcp_server/http.py" = ["F821"]
"mempalace/mcp_server/runtime.py" = ["F821"]
# cli fragments exec into the package namespace. Names they use live in
# __init__.py and earlier fragments, so F821 is expected; __init__.py
# imports those names for the fragments, so F401 is expected there too.
"mempalace/cli/__init__.py" = ["F401"]
"mempalace/cli/_*.py" = ["F821"]
"mempalace/cli/cmd_*.py" = ["F821"]
"mempalace/cli/parser.py" = ["F821"]
# searcher fragments exec into the package namespace.
"mempalace/searcher/__init__.py" = ["F401"]
"mempalace/searcher/*.py" = ["F821"]
# palace fragments exec into the package namespace.
"mempalace/palace/__init__.py" = ["F401"]
"mempalace/palace/*.py" = ["F821"]
# pytest idioms: `pytest.raises(Exception)` around backends that raise their
# own driver errors, and bare attribute reads inside `pytest.raises` blocks.
"tests/**" = ["B017", "B018"]
# The comment blocks document transcript formats with sample JSON lines.
"tests/test_normalize.py" = ["ERA001"]

[tool.ruff.lint.mccabe]
max-complexity = 25

[tool.ruff.format]
quote-style = "double"

[tool.pytest.ini_options]
testpaths = ["tests"]
pythonpath = ["."]
addopts = "-m 'not benchmark and not slow and not stress'"
markers = [
    "benchmark: scale/performance benchmark tests",
    "slow: tests that take more than 30 seconds",
    "stress: destructive scale tests (100K+ drawers)",
]
filterwarnings = [
    # Many tests build raw-chromadb palaces directly (no recorded embedder
    # identity), which correctly emits this on open. The behavior itself is
    # asserted in tests/test_embedder_identity.py; ignore the fixture noise.
    # Matched by message (not by class path) so pytest does not import the
    # mempalace package at config time — importing it before coverage starts
    # would drop module-level lines from the report.
    "ignore:palace collection.*has no recorded embedder identity",
]

[tool.coverage.run]
source = ["mempalace"]
# pymilvus ships in the milvus extra, not in dev. Counting this module in the
# default total (it sits near 20% without the extra) pulls the package gate
# under its threshold. The test-milvus job measures it instead.
omit = [
    "*/mempalace/backends/milvus.py",
]

[tool.coverage.report]
fail_under = 85
# Match the two decimals pytest-cov prints. Precision 0 would round 84.59 up
# to 85 and the gate would pass.
precision = 2
show_missing = true
exclude_lines = [
    "if __name__",
    "pragma: no cover",
]

[tool.mempalace.coverage]
# Windows gate (ChromaDB file-lock cleanup). tests/coverage_gate.py applies
# this over fail_under when sys.platform == "win32". Local pytest and CI both
# read it; neither passes --cov-fail-under.
fail_under_windows = 80