cekura-predefined-metrics

Enabling a predefined metric requires two steps: toggle it on at the project level for simulation or observability, then attach it to individual evaluators by its code. Missing either step means the metric appears available but never fires.

Free metrics include Expected Outcome, Tool Call Success, Mock Tool Call Accuracy, Latency, Detect Silence, Infrastructure Issues, Interruption Score, Repetition Score, Average Pitch, Talk Ratio, and Words Per Minute. Most others cost 0.2 credits per call, while Hallucination costs 0.6 credits.

Detect Silence returns False when neither speaker talks for longer than the configured threshold, measuring mutual silence. Infrastructure Issues measures only main-agent silence, isolating agent-side outages from natural conversation pauses.

For observability calls, AI Interrupting User, User Interrupting AI, Talk Ratio, and Gibberish Detection require stereo recordings with separate speaker channels. Transcription Accuracy and Gibberish Detection also require audio data and cannot run on text-only calls.

Common causes include missing the project-level toggle or evaluator attachment, omitting required configuration keys like dropoff_nodes or topic_nodes, or using a metric outside its supported mode such as Transcription Accuracy on observability calls.

Use custom metrics when evaluating business-specific workflows, checking behavior against your own system prompt, or combining multiple signals into one score. Predefined metrics cover common quality dimensions with no prompt engineering required.