Features

  • Tracing — nested traces/spans for each LLM/STT/TTS step
  • Usage & cost — per-call token usage and cost roll-ups
  • Latency — span timings, TTFT / completion-start
  • Sessions — group multi-turn interactions
  • Datasets — build datasets from real traces for evaluation
  • Scores / evals — attach LLM-as-judge or programmatic scores to traces
  • Prompt management — versioned prompts
  • OpenTelemetry — OTEL-compatible ingestion

Gotchas

  • Shared OTEL exporter + framework auto-instrumentation can create orphan/duplicate traces — isolate or filter exporters.
  • Float usage_details may be dropped — send integers (e.g. audio_duration_ms).
  • Spans that end at async drain look “instant” — backdate with real duration_ms / TTFT.
  • Tag provider as a property so dashboards group by backend.
  • Trace sims + evals, not only live calls.

Connections