Summary

Migrates Python packaging and dependency management from setup.py + requirements*.txt to a standard PEP 621 pyproject.toml managed with uv, adding a uv.lock for reproducible installs across dev, CI, and Docker.

Key points:

• End users are unaffected: the published wheel is standard, so pip install dashai keeps working exactly as before.
• The CPU/CUDA requirements variants are now extras: uv sync --extra cpu / --extra cuda pick the right PyTorch index automatically via [tool.uv.sources] — this solves the torch-index problem that made us abandon pip-compile in the past (see the old note in requirements.txt).
• CI now installs from the lockfile (uv sync --locked), so builds are reproducible and dependency caching is handled by setup-uv.
• Fixes a packaging bug: the published wheel included the whole tests/ package (find_packages() had no filter); packages are now scoped to DashAI*.

Type of Change

• Backend change
• Frontend change
• CI / Workflow change
• Build / Packaging change
• Bug fix
• Documentation

Changes (by file)

• pyproject.toml: full PEP 621 metadata (moved from setup.p requirements*.txt), cpu/cuda extras with per-extraPyTorch/llama-cpp indexes, dev dependency group. Setuptools reANIFEST.in keeps bundling the frontend.
• uv.lock (new): universal lockfile (288 packages, all platf
• setup.py, requirements{,-dev,-cpu,-cuda,-plugins}.txt: d.txt` was unreferenced and pinned torch 1.13 from 2022).
• .github/workflows/build-test.yaml, db-migrations.yaml: `` + uv run ....- `.github/workflows/publish.yml`: `uv build` + `uvx twine`; P--extra cpu --no-dev` and `uv run --no-sync`; release version isnow extracted from `pyproject.toml`.- `.github/workflows/docs.yaml`, `pre-commit.yaml`: install inpip install --system` (the Docusaurus build and a `language:system` hook invoke `python` directly).
• Dockerfile, Dockerfile.cuda: install with uv sync --locDA image builds llama-cpp with CMAKE_ARGS="-DGGML_CUDA=on"` as
  before.
• update-server.sh: uses uv when available, falls back to pi
• README.rst, docs/**/dev-setup.md (en/es), CLAUDE.md, uv is the recommended path (including uv tool install dashai`),
  pip instructions kept.

Testing

• Full backend suite passes with the new environment: uv run pytest tests/ → 541 passed. - Wheel built with uv build passes twine check; verified i, DashAI/alembic, seeds and static images, and no longerships tests/`.
• Smoke test: server boots from the uv-managed venv and `/api/mponent registry.
• uv sync --locked --extra cpu and --extra cuda resolve coend locally; the CUDA llama-cpp compilation is only exercised by
  Dockerfile.cuda, same as before).

Notes

• pip survives on purpose in three places — please don't " installer (DashAI/back/plugins/utils.py) runs in the enduser's env where pip is guaranteed; the AppImage job uses pythupdate-server.sh` falls back to pip if the server lacks uv.
• The PyTorch index selection in the extras lives in [tool.uval and not published to PyPI — pip users still follow the manual --index-url` instructions in the README (unchanged).
• Gotcha for contributors: uv run re-syncs the env to the deuv sync --extra cpu, use uv run --no-sync or re-run thesync.
• Contributors now need uv installed (one-liner, see README); e watched on its first run against production.