Metrics

Aggregate views over traces — latency, errors, cost, feedback.

The Metrics page shows rolling aggregations over the time window selected in the top-right picker. Every chart is interactive — click a point to drill into the underlying traces.

Charts

Volume

Trace count per bucket. First thing to check after a deploy — sudden drops mean the ingest is broken; sudden spikes mean something looped.

Error rate

Fraction of traces with setError() or uncaught exceptions. Dashed overlay shows the previous period for comparison.

Latency percentiles

p50, p95, p99 of trace wall-clock duration. Toggle by metric to overlay. Most teams treat p95 as the SLI.

Cost

Sum of (prompt_tokens + completion_tokens) × model_price per bucket. Split by model on click to see which model dominates.

Feedback positive rate

Fraction of feedback-labeled traces with label="good". Low-signal if feedback volume is low — watch the companion count chart.

Eval pass rate

Fraction of pass labels over the eval runs in the window. Per-evaluator breakdown via the metric picker.

Filters

Apply across all charts on the page:

• Project
• Environment (production / staging / your custom labels)
• Model
• Metadata key/value — e.g. metadata.tier = "pro" or metadata.feature = "rag_v2"

Filters persist in the URL, so you can bookmark a view.

Drilling down

• Click any point on a chart → jump to Traces filtered by that bucket and current filters
• Click a model label in the legend → filter all charts to that model
• Right-click a chart → Copy embed URL for sharing with teammates

Anomaly detection

Anomaly detection is available on all plans, including Starter.

Every chart has an optional anomaly overlay (enable via the Anomalies toggle in the top-right). TruLayer fits a baseline on the trailing 7 days, bucketed by hour-of-week, and flags points more than 3σ outside the band. Anomalies appear as red dots on the chart.

Click any anomaly dot to jump to the Traces page filtered to that time bucket with the originating filter set applied — useful for going from “latency p95 spiked at 14:05” to the exact 12 slow traces behind it.

Anomalies also drive the failure-rate rules configured on the Failures page — see Failures → Alert rules for how to subscribe.

Via API

curl "https://api.trulayer.ai/v1/metrics?project_id=rag-app&from=2026-04-19T00:00:00Z&to=2026-04-19T23:59:59Z" \
  -H "Authorization: Bearer $TRULAYER_API_KEY"

Returns latency percentiles, error rate, total cost, trace count. See API reference.