adk_eval_llm_judge (erlang_adk v0.7.0)

View Source

Bounded first-party rubric judge backed by an ADK LLM provider.

The judge is an explicit full-case metric. It is never selected implicitly by adk_eval_set. Each invocation runs in the caller's existing evaluation sample worker and owns one monitored provider worker; there is no global judge process or serialization point.

Summary

Types

judge_error/0

-type judge_error() :: {llm_judge, atom()}.

Functions

score_case(EvalInput, Config)

-spec score_case(map(), map()) -> {ok, number(), map()} | {error, judge_error()}.

validate_config(Config)

-spec validate_config(map()) -> ok | {error, judge_error()}.