pyproject.toml

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

[project]
name = "boldt-embed-de"
version = "0.1.0"
description = "German-first embedding model family (causal, bidirectional, reranker) based on Boldt/Boldt-DC-350M."
readme = "README.md"
requires-python = ">=3.10"
license = { text = "Apache-2.0" }
authors = [{ name = "Boldt-Embed-DE maintainers" }]
keywords = ["embeddings", "german", "retrieval", "reranker", "matryoshka", "sentence-transformers"]

# The core package and ALL validation gates run on the Python standard library
# only. Heavy ML dependencies are optional and used solely by the training and
# real-benchmark code paths (which are dry-runnable without weights here).
dependencies = []

[project.optional-dependencies]
train = [
  "torch>=2.1",
  # Qwen3 teacher model cards require a recent Transformers; <4.51 cannot load Qwen3.
  "transformers>=4.51.0",
  # 3.x ships CachedMultipleNegativesRankingLoss / (Cached)GISTEmbedLoss / MatryoshkaLoss
  # and Qwen3 CrossEncoder support used by the 2026 teacher/student workflow.
  "sentence-transformers>=3.0",
  "accelerate>=0.30",
  "datasets>=2.19",
  "peft>=0.11",
]
eval = [
  "mteb>=1.12",
  "sentence-transformers>=3.0",
]
# flash-attn is intentionally NOT a hard dependency: its wheels need a matching CUDA
# toolchain + torch already installed and frequently fail under build isolation. Install
# it manually if desired (the teacher loader falls back to eager attention when absent):
#   pip install flash-attn --no-build-isolation
flash = [
  "flash-attn>=2.5.0",
]
dev = [
  "ruff>=0.5",
]

[project.scripts]
boldt-validate = "boldt_embed.cli:main_validate"
boldt-smoke = "boldt_embed.cli:main_smoke"
boldt-bench = "boldt_embed.cli:main_bench"

[tool.setuptools.packages.find]
where = ["src"]

[tool.ruff]
line-length = 100
target-version = "py310"