Observability
A deeper, platform-admin view of call quality — how fast the agent responds, where errors happen, and which calls need a look. It’s a debugging and quality tool, separate from the everyday Call history.
Where: Observability (visible to super-admins).
Fleet overview
Across recent calls you see:
- Volume — calls in the last period and how many are active now.
- Error rate — share of calls with errors.
- Response times — median (p50) and worst (p95) latency.
- Calls that need attention — sorted by errors or slow turns, so you triage the worst first.
- Top issues — the most common error/warning types.
The latency waterfall
Open any call to see a per-turn breakdown of where each response’s time went:
| Segment | What it is |
|---|---|
| Wait | Endpointing — how long the agent waited to be sure the caller finished speaking. |
| STT | Turning speech into text. |
| LLM | The model thinking (time to first token) — usually the biggest piece. |
| Action | Time spent in a booking/API/SMS action — shown as a concurrent bar (it overlaps the reply, so it isn’t double-counted). |
| TTS | Turning the reply into speech. |
This makes it obvious why a turn was slow — e.g. a slow model turn vs. a slow calendar lookup.
Errors & recordings
Each call also surfaces any errors or transport issues, the transcript, and the recording (if enabled) — everything you need to understand a bad call without guesswork.
Use Observability to answer “why did that call feel slow / go wrong?” — it’s for tuning and debugging, not day-to-day operation.
Last updated on