Documentation

Analytics

Review latency, token usage, and execution trends.

Overview

The Analytics view aggregates trace data across all projects and environments into summary metrics, time-series charts, and breakdowns. Data is fetched from the GET /analytics endpoint, which returns a pre-computed response including summary stats, hourly chart points, and dimension breakdowns.

Summary Metrics

The summary section displays eight key metrics computed from all available traces:

Metric	Description	Source
Total traces	Count of all trace documents	collection.count_documents()
Success rate	Percentage of traces with status success	traces with status == success / total
Average latency	Mean wall-clock time across all traces	sum(latencies) / total
P95 latency	95th percentile latency	Sorted latencies, 95th percentile index
Total token usage	Sum of token_count across all traces	sum(trace.token_count)
Failed traces	Count of traces with status failed	count where status == failed
Warning traces	Count of traces with status warning	count where status == warning
Slow requests	Count of traces with latency >= 1500ms	count where slow_request == true

Time-Series Charts

Traces from the last 24 hours are bucketed by hour. Each bucket provides three data points per hourly window:

Average latency — mean of all trace latencies in that hour
Token usage — sum of token counts in that hour
Trace count — number of traces recorded in that hour

Buckets are computed by grouping traces on their created_at field formatted to HH:00. Traces older than 24 hours are excluded from the chart data.

Breakdowns

The analytics endpoint returns four breakdown dimensions:

Breakdown	Key Used	Example
Status breakdown	trace.status	success: 42, failed: 3, warning: 5
Model breakdown	trace.model_name or unknown	gpt-4o: 20, gpt-4.1-mini: 30
Project breakdown	trace.project_name or project_id	default: 40, search-service: 10

Info

Breakdowns are sorted by count (most common first). The API processes up to 5000 documents to compute analytics. For larger datasets, filter by project or time range.

Recent Failures

The analytics view also surfaces the five most recent traces that have issues — failed traces, warning traces, slow requests, or traces with retries. This provides a quick overview of system health without navigating to the Failures view.