Skip to content

release-notes-v0.0.7.md

Latest commit

 

History

History
27 lines (20 loc) · 2.04 KB

release-notes-v0.0.7.md

File metadata and controls

27 lines (20 loc) · 2.04 KB
Raw

umbrela v0.0.7

Initial release-note scaffold for the packaged CLI and prompt-template workflow.

Included In This Baseline

  • The supported interpreter floor is now Python 3.12 across package metadata, .python-version, contributor setup docs, and GitHub Actions validation.
  • Packaged umbrela CLI for direct judging, qrel-backed evaluation, schema inspection, validation, doctor, and view commands.
  • FastAPI umbrela serve command exposing GET /healthz and POST /v1/judge on port 8084 by default.
  • Direct judge input now also accepts Anserini REST candidates where candidates[].doc is a plain string, so Anserini search results can be piped directly into POST /v1/judge without a jq reshape step.
  • Direct judge input now also accepts single-record castorini.cli.v1 envelopes from upstream tools such as rank_llm, so search | rerank | judge can be piped through POST /v1/judge without unwrapping .artifacts[0].value[0] first.
  • Direct umbrela judge --output json and POST /v1/judge responses are now compact by default. The default judgment records include query, passage, and judgment, with reasoning remaining opt-in via --include-reasoning. The older verbose fields prediction, result_status, and prompt are now exposed only through the new --include-trace flag, with --redact-prompts available to suppress raw prompt text.
  • YAML-backed prompt templates with the legacy flat-prompt rendering contract preserved.
  • Offline-first contributor workflow built around uv.
  • Explicit uv lock --check enforcement in the local quality gate and GitHub Actions so committed uv.lock stays aligned with contributor and CI environments.

Migration Notes

This baseline establishes the release-note policy for future changes.

Document a migration note whenever a change affects:

  • prompt semantics or rendered prompt structure
  • parsing heuristics or judgment extraction
  • backend defaults or provider selection behavior
  • qrel, evaluation, or artifact output formats
  • supported Python version or default contributor runtime