发布

  • [OPIK-4491] [SDK] [DOCS] fix: optimizer multi-metric span wiring, fail-fast scoring, and examples (#5276)

    frostbyte_neo 发布于 2026-02-24 13:40:37 +00:00

    • feat(sdk): add LevenshteinAccuracyMetric helper

    • feat(sdk): normalize span cost and duration with target

    • fix(sdk): propagate task_span and fail on objective metric errors

    • docs(sdk): clarify multi-metric normalization and warn on out-of-range values

    • chore(sdk): centralize noisy HF/transformers warning filters

    • docs(sdk): add hrpo cost-duration-accuracy example and docs link

    • feat(sdk): add explicit invert controls for span efficiency metrics

    • fix(sdk): address reviewer feedback for span metric docs and compatibility

    • docs(sdk): add missing docstrings for evaluation helpers

    • docs(sdk): default example scoring to validation split

    • fix(python-sdk): avoid KeyError on orphan local-recording span links

    • fix(python-sdk): process missing create retries in local emulator

    • fix(sdk): preserve metric API compatibility and forward task_span

    • refactor(sdk): normalize span metric naming and remove duration alias arg

    • docs(sdk): document missing-timestamp error behavior for duration metric

    • docs: Update define_metrics.mdx

    • refactor(optimizer): unify span metrics naming and centralize warning filters

    • Update apps/opik-documentation/documentation/fern/docs/agent_optimization/optimization/define_metrics.mdx

    Co-authored-by: baz-reviewer[bot] <174234987+baz-reviewer[bot]@users.noreply.github.com>

    • fix(metrics): raise on missing task_span for span metrics

    • docs(metrics): clarify SpanCost reason in normalized example

    • fix(metrics): default span metrics to soft-fail on missing task_span


    Co-authored-by: baz-reviewer[bot] <174234987+baz-reviewer[bot]@users.noreply.github.com>

    下载附件