Skip to main content

Metrics

The Metrics page provides aggregated, macro-level performance and health charts for your organization's AI infrastructure. It transforms raw ClickHouse log streams into real-time analytical dashboards for trend monitoring and capacity planning.

Navigate to OBSERVABILITY → Metrics in the Aegis Console to inspect system-wide metric dashboards.

KPI Cards Summary​

Four key performance indicators appear across the top of the Metrics panel:

  • Total Traces: Total execution count during the time window.
  • Fail Rate: Overall failure percentage across all operations.
  • Avg Latency: Average response time in milliseconds (ms).
  • Total Tokens: Aggregate token consumption across all model calls.

Visual Metric Dashboards​

The Metrics dashboard features four interactive analytical charts:

1. Requests & Errors Over Time​

  • Description: Dual-line time series chart displaying total request throughput (Requests) against error frequency (Errors).
  • Use Case: Quickly identify traffic spikes, service outages, or sudden clusters of HTTP 429 / 500 errors.

2. Latency Trend (Avg ms)​

  • Description: Line trend chart illustrating average response times in milliseconds over time.
  • Use Case: Monitor upstream model provider latency (e.g., OpenAI or Vertex AI responsiveness) and detect performance degradation.

3. Token Consumption (Prompt vs Completion)​

  • Description: Stacked bar chart highlighting Prompt Tokens vs Completion Tokens per time bucket.
  • Use Case: Analyze input prompt overhead versus generated output volume to optimize prompt engineering and model pricing strategies.

4. Top Agents by Calls​

  • Description: Horizontal bar chart ranking agents by total invocation count across the organization.
  • Use Case: Identify the most active AI agents and allocate quota or compute resources accordingly.

Filtering by Agent​

To inspect metrics scoped to a single agent:

  • Append ?agent=<agent_id> to the URL or navigate directly to Agents → Agent Details → Metrics tab.
  • In agent-scoped view, the Top Agents chart is automatically hidden to focus on the single agent's request, latency, and token trends.