Castorini Onboard

Castorini Onboard

Unified development environment setup for the Castorini Python repos, whether the user wants one repo or several.

Supported Repos

Context Detection

1. If cwd is inside one of the 4 repos (check for pyproject.toml with matching project name) → install that one.
2. If the user explicitly asks for one repo, install only that repo.
3. If cwd is the monorepo root or elsewhere and the request is ambiguous → ask which repo(s) to install, or install all if the user says so.

Prerequisites

• Python 3.11+
• Git (SSH access to github.com:castorini)

uv Detection

command -v uv

If present, use uv silently. If absent, ask the user once: install uv or proceed with pip.

Virtual Environment Preference

• Prefer a shared .venv-shared in the current workspace root over repo-local .venv directories.
• Before creating any environment, check whether .venv-shared already exists in the current workspace root and reuse it if present.
• Only fall back to a repo-local environment if the shared environment is unavailable or the user explicitly asks for isolation.

Install Flow

1. Resolve which repository or repositories are in scope.
2. Clone any missing repositories with SSH access to castorini/<repo>.
3. Reuse .venv-shared when it already exists in the workspace root.
4. Prefer uv for sync and dependency-group support.
5. Fall back to pip only when uv is unavailable or the user explicitly asks.
6. For the common multi-repo cloud/API source install, prefer one shared editable install command:

uv pip install \
  -e './ragnarok[cloud,api]' \
  -e './nuggetizer[api]' \
  -e './umbrela[cloud,api]' \
  -e './rank_llm[openai,api]'

This pulls OpenAI support for all four repos and API or uvicorn support for each repo from its own declared extras instead of relying on an indirect transitive install from another package.

7. Run the repo CLI doctor --output json smoke test after install.
8. Run pre-commit install in each repository after a source install.

Post-Install (all source installs)

pre-commit install

Reference Files

• references/extras.md — Per-repo optional dependency stacks
• references/install-recipes.md — Per-repo clone, install, and smoke-test command sequences

Gotchas

• uv sync --group dev understands dependency groups; pip install -e . does not. If you fall back to pip, install dev tools manually.
• ragnarok uses the package name pyragnarok on PyPI even though the repo and CLI command are ragnarok.
• rank_llm uses the package name and CLI binary rank-llm, while the repository and import package are spelled rank_llm.
• rank_llm supports openai and api as separate extras; use both when the user wants OpenAI-backed reranking plus HTTP serving helpers.
• umbrela only needs Java 21 for pyserini evaluation workflows, not for the default cloud-oriented development install.
• When reusing .venv-shared, make sure it was created with a Python version compatible with the target repo instead of blindly reusing an older interpreter.
• Run smoke tests from inside the target repository so editable installs and local entry points resolve correctly.