Skip to content

Feedback

autogen.beta.eval._types.Feedback `dataclass` #

Feedback(key, score=None, value=None, comment=None, detail=None)

A single piece of feedback produced by a scorer.

Exactly one of score or value is typically populated — score for numeric / boolean grades, value for categorical labels. Both being None is valid and represents a "no signal" feedback (e.g. a scorer that crashed mid-evaluation).

PARAMETER	DESCRIPTION
`key`	Stable identifier for this feedback, usually the scorer's function name. Pass rates and stats on :class:`~autogen.beta.eval.RunResult` are looked up by this key. TYPE: `str`
`score`	Numeric or boolean grade. TYPE: `ScoreValue` DEFAULT: `None`
`value`	Categorical label, used for slicing aggregates. TYPE: `ValueLabel` DEFAULT: `None`
`comment`	Free-form human-readable explanation. Surfaces in run JSON; useful for LLM-as-judge rationales or scorer error traces. TYPE: `str \| None` DEFAULT: `None`
`detail`	Optional structured evidence behind this feedback — a JSON-safe mapping serialized into the run JSON for programmatic access (e.g. a failure attribution's decisive step / responsible agent, or a judge's per-order swap verdicts). Supplementary only: aggregation never reads `detail` — `score` / `value` remain the graded signal. Typed at the source (producers serialize a typed model into it), so it is evidence, not a grab-bag. TYPE: `dict[str, Any] \| None` DEFAULT: `None`

key `instance-attribute` #

key

score `class-attribute` `instance-attribute` #

score = None

value `class-attribute` `instance-attribute` #

value = None

comment `class-attribute` `instance-attribute` #

comment = None

detail `class-attribute` `instance-attribute` #

detail = field(default=None, compare=False)