Skip to content
Arbitex Docs

Model quality

What it does

Section titled “What it does”

Model quality is analytics and LLMOps-style observability for your org’s traffic: drift, anomaly, and fairness-style quality lenses over time. It shows an average-score tile, anomaly and evaluation-volume counts, a quality trend (score and anomalies over time), average score by lens, evaluation volume by model, fired quality-alert history, and the real per-score alert firing thresholds.

Configure in the UI

Section titled “Configure in the UI”
The Model quality panel showing summary tiles for average score and anomalies, a quality trend line chart, a score-by-lens bar chart, and an evaluations-by-model donut.
Monitoring & Reporting → Model qualityCitadel synthetic data on a non-production Arbitex demo instance. Never client data.

  1. Open Monitoring & Reporting → Model quality. Pick a Period (Day / Week / Month) for the trend chart’s granularity.

  2. Review the summary tiles — Average score, Anomalies, Evaluations, Lenses tracked. Only the generic score/anomaly/volume aggregates show here; the four named per-score thresholds (for example hallucination, relevance, toxicity, groundedness) appear live in Alert configuration below.

  3. Review Average score by lens and Evaluations by model.

  4. Review Quality alert history (real fired alerts: score, threshold, observed value, model, fired-at) and Alert configuration (the real per-score firing thresholds). Both read “Temporarily unavailable” if the platform read can’t be reached — never a fabricated event or threshold.

  5. Scroll to the retained Recent quality events table (lens, metric, model, value, window, anomalous flag).

  6. Export report (PDF/CSV) is a disclosed-pending affordance — the buttons are visible but disabled; there’s no backing export endpoint yet.

API equivalent

Section titled “API equivalent”
Terminal window
curl -X POST "https://admin.arbitex.ai/api/v1/admin/monitoring/quality/timeseries" \
  -H "Authorization: Bearer $ADMIN_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"bucket": "day", "series": ["score", "anomalies", "points"]}'

The lens and model breakdowns read quality/breakdown; the raw drill-down table reads quality/list.

Plan & availability

Section titled “Plan & availability”

Available on every plan.

Deployment notes

Section titled “Deployment notes”

Alert history and alert configuration are cross-plane platform reads; in air-gapped deployments without a path to the platform, they show “temporarily unavailable” rather than being hidden.

Section titled “Related”