Analytics

The Analytics page (/analytics) provides aggregate views of your trace and span data. Use it to track costs, identify latency regressions, compare model usage, and monitor request volume over time.

Built-in charts

The dashboard ships with five default charts. Each chart respects the active time range and any filters you've applied.

A line chart showing total estimated cost per time bucket (hour or day, depending on the selected range). Hover over a point to see the exact cost for that period. The chart includes a cumulative total line so you can track spend against a budget.

Latency percentiles

A multi-line chart showing P50, P95, and P99 latency for LLM call spans. This is the most useful chart for spotting performance regressions — if P99 spikes while P50 stays flat, you have a tail-latency problem.

Percentile	What it tells you
P50	Median latency — what most requests experience
P95	95th percentile — worst case for most users
P99	99th percentile — tail latency, often caused by cold starts or rate limits

Token usage by model

A stacked bar chart showing input and output tokens broken down by model. Each bar represents a time bucket, and the segments show which models consumed the most tokens. Useful for spotting unexpected model usage — for example, if a fallback model is getting more traffic than expected.

Request volume

A bar chart showing the number of traces created per time bucket. Overlays show the breakdown by status (completed, failed, running). A sudden drop in volume can indicate an upstream issue; a spike in failures warrants investigation.

Error rate

A line chart showing the percentage of spans with failed status over time. The chart plots both the raw error count and the error rate (failed / total). A rising error rate is the most actionable signal on the dashboard.

Time range selection

The time range picker in the top-right corner controls all charts. Available options:

Option	Range
Last 1 hour	Rolling 1h window
Last 24 hours	Rolling 24h window
Last 7 days	Rolling 7d window
Last 30 days	Rolling 30d window
Custom	Date picker for arbitrary start and end

Charts automatically adjust their bucket size based on the range — hourly buckets for ranges under 7 days, daily buckets for longer ranges.

Per-model breakdowns

Click any model name in the Token Usage chart to drill into a model-specific view. This shows:

Latency distribution — Histogram of response times for that model only
Cost trend — Cost over time for that model
Token efficiency — Average output tokens per request, useful for detecting prompt bloat
Error rate — Failure percentage for that model

Use this to compare models side by side. If you're evaluating a switch from gpt-4o to claude-3-5-sonnet, the per-model view shows exactly how each performs in your production traffic.

Filters

Analytics support the same filter syntax as the Traces and Spans pages. Type filters into the search bar to narrow the data:

model:gpt-4o provider:openai

kind:llm_call status:completed cost:>0.01

Filters apply to all charts simultaneously. This is useful for isolating specific workloads — for example, filtering to name:summarize to see cost and latency for your summarization pipeline only.

Custom charts

Create custom charts by clicking New chart in the top-right. Each custom chart requires:

Metric — What to measure: cost, latency, tokens, count, or error_rate
Group by — How to break down the data: model, provider, kind, name, or none
Aggregation — How to combine values in each bucket: sum, avg, p50, p95, p99, min, max
Chart type — line, bar, or stacked_bar

Custom charts are saved per-project. Any team member can see and edit them.

Example: cost per user

To track cost per user, tag your traces with a user ID in metadata, then create a custom chart:

Metric: cost
Group by: metadata.user_id
Aggregation: sum
Chart type: stacked_bar

Exporting data

Export analytics data for use in external tools:

CSV export — Click the download icon on any chart to export the underlying data as CSV. Each row contains the time bucket, metric value, and group-by dimension.
API access — Query analytics programmatically:

# Summary stats for the current project
curl "https://api.traceway.ai/api/analytics/summary" \
  -H "Authorization: Bearer tw_sk_..."

# Detailed time-series data
curl -X POST "https://api.traceway.ai/api/analytics" \
  -H "Authorization: Bearer tw_sk_..." \
  -H "Content-Type: application/json" \
  -d '{
    "metric": "cost",
    "group_by": "model",
    "aggregation": "sum",
    "since": "2024-06-01T00:00:00Z",
    "until": "2024-06-30T23:59:59Z",
    "bucket": "day"
  }'

The response contains an array of time-series points, each with a timestamp, value, and optional group label.

On this page