use uv for dependency management and running python
To run the evals (search eval, or the RAG end-to-end eval, plain or Opik-tracked), see docs/evals.md.
periodically commit to git. use double quotes for commit messages, and put the entire message in one line.
Recorded eval results must describe the suite as it stands now. When cases are added or changed, re-run the eval and record the new number. Never publish a score qualified as predating some of the cases ("before case #N was added", "before these cases existed") — re-run and replace the number instead.