Skip to main content

​
Observability Hub

The Observability Hub aggregates the execution telemetry your agents emit via the Fact0 SDK (or OpenTelemetry) into four views: LLM Calls, Tool Executions, Error Analytics, and Conversations. All metrics are scoped to the last 30 days by default.
Metrics populate from execution spans. If you haven’t integrated the SDK yet, start with the Quickstart or the SDK overview.

​
How data gets here

Every time your agent completes a model call or tool invocation, the SDK emits a span to Fact0. The Observability Hub reads those spans and aggregates them server-side: 
import fact0, os

client = fact0.Client(api_key=os.environ["FACT0_API_KEY"])

with client.telemetry.execution(agent_id="my-agent") as ex:

    # This span produces an LLM Calls data point
    with ex.span("gpt-4o call", span_type="MODEL_INVOCATION") as s:
        s.complete(model_invocation={
            "model_name": "gpt-4o",
            "model_provider": "openai",
            "prompt_tokens": 512,
            "completion_tokens": 128,
            "latency_ms": 870,
        })

    # This span produces a Tool Executions data point
    with ex.span("search", span_type="TOOL_CALL") as s:
        s.complete(tool_call={"tool_name": "web_search", "duration_ms": 340})
If you use OpenTelemetry, gen_ai.* semantic convention attributes are automatically classified as MODEL_INVOCATION and tool spans as TOOL_CALL — no code changes needed. See OpenTelemetry integration.

​
LLM Calls tab

Aggregated statistics for all MODEL_INVOCATION spans over the selected period.
MetricDescription
Total callsRaw span count
Success rateSpans without error / total
Avg / P50 / P95 / P99 latencyPercentile latency in ms
Prompt tokensInput tokens consumed
Completion tokensOutput tokens generated
Estimated cost (USD)Calculated from provider pricing tables
By model — a per-model breakdown of call count, token usage, latency, error count, and estimated cost across all providers your agents use. Call frequency chart — hourly/daily bucketed call volume over time, useful for spotting traffic spikes or agent regressions.

​
Tool Executions tab

Aggregated statistics for all TOOL_CALL spans.
MetricDescription
Total callsRaw span count
Success rateSpans without error / total
Avg durationMean execution time in ms
By tool — per-tool breakdown of call count, duration, error count, and success rate. Useful for identifying slow or unreliable tools.

​
Error Analytics tab

Error frequency derived from spans with has_error = true.
  • By model — which LLM providers are producing errors and at what rate
  • By tool — which tools fail most often
  • Time series — error count over time to track regressions or incidents

​
Conversations tab

Conversations groups related agent executions into sessions — a higher-level view of multi-turn agent interactions with a user or system. A session is identified by a session_id metadata field you pass when starting an execution: 
with client.telemetry.execution(
    agent_id="support-bot",
    metadata={"session_id": "sess_abc123"},
) as ex:
    ...
Each session card shows:
  • Agent name and session ID
  • Turn count (number of executions in the session)
  • Total tokens consumed and estimated cost
  • Session status (active / completed)
Selecting a session opens a turn trace view showing each execution in sequence, with its LLM calls, tool calls, token counts, latency, and cost.

​
Prompt Registry

The Prompt Registry stores versioned prompt templates your agents fetch at runtime, keeping prompts auditable and out of application code. See the Prompt Registry guide for full details.

​
Date range

All tabs respect a shared time range (default: last 30 days). Pass from and to as ISO 8601 timestamps if querying the API directly: 
GET /v1/me/analytics/llm?from=2026-06-01T00:00:00Z&to=2026-06-30T23:59:59Z