pyproject.toml

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

[tool.setuptools.packages.find]
include = ["backend*"]

[tool.ruff]
line-length = 120

[project]
name = "embeddorium"
version = "0.1.0"
requires-python = ">=3.11"

# Shared by every actor
dependencies = [
  "pydantic>=2",
  "sqlalchemy>=2",
  "psycopg2-binary>=2.9",
  "dramatiq[rabbitmq,watch]>=1.17",
  "python-dotenv>=1",
]

[project.optional-dependencies]

# worker-web-source-processor
web = [
  "requests>=2",
  "trafilatura>=2",
  "langchain-text-splitters>=0.3",
]

# worker-embed-chunks (in docker-compose): the embedding providers are all
# remote/API (ollama, openai) or the trivial in-process mock, so the container
# stays lightweight — no torch/onnxruntime/sentence-transformers. langchain-
# ollama/ollama and the openai client are HTTP clients (httpx-based).
embed = [
  "qdrant-client>=1",
  "langchain-ollama>=0.3",
  "httpx>=0.27",
]

# Embeddings tester API (server/): FastAPI service that embeds source/candidate
# texts and scores them against each other. Reuses backend's VectorStore +
# OllamaEmbedClient, so it needs the same vector/embed deps plus the web stack.
# The actor-config endpoint imports plugin modules to build their registries.
# Keep their import-time dependencies here: chunkers need
# langchain-text-splitters, and parse_source needs trafilatura through the HTML
# parser. Missing either dependency silently removes the affected strategies.
server = [
  "fastapi>=0.103",
  "uvicorn>=0.23",
  "qdrant-client>=1",
  "langchain-ollama>=0.3",
  "langchain-text-splitters>=0.3",
  "trafilatura>=2",
  "numpy>=1",
  # Vector-space projection (/projection): PCA runs on numpy alone; MDS,
  # Isomap, Spectral and t-SNE all come from sklearn.manifold. Optional at
  # runtime — the endpoint reports an algorithm as unavailable rather than
  # failing — but shipped by default so the Inspect tab offers a real choice of
  # layouts out of the box.
  "scikit-learn>=1.4",
  "httpx>=0.27",
]

# UMAP for /projection. Kept out of `server` because umap-learn drags in
# numba/llvmlite — a few hundred MB of image for one more layout. Install this
# extra to unlock it; without it the endpoint reports UMAP as unavailable and
# the UI disables the option with that reason, so nothing breaks either way.
umap = [
  "umap-learn>=0.5",
]

mcp = [
  "fastmcp",
  "httpx>=0.27",
]

agent = [
  "langgraph>=1.1",
  "langchain-mcp-adapters>=0.2",
  "langchain-openai>=0.3",
  "langchain-ollama>=0.3",
  "python-dotenv>=1",
]

dev = [
  "pytest>=8",
  "ruff>=0.8",
  "mypy>=1",
  "watchdog>=6",
]

[project.scripts]
agent-generate = "backend.agent.generate:main"