Operations

Sessions & analytics

Every conversation, in full — plus where your usage is actually going.

Once characters are live, most of your time in the panel is spent here. Sessions tell you what was said and why an answer went wrong; Analytics tells you which character is busy and which model is eating the allowance.

Sessions — the transcripts

Sessions list showing recent conversations with speaker, status, start time, turn count and cost
Every conversation across every character, newest first.

Each row is one conversation — a website visitor, a kiosk user, a phone caller or a live-stream viewer. Open means it is happening right now; Closed means it finished.

Opening a session gives you the whole exchange: what the visitor said, what the character replied, when, how many tokens each turn used, and — where audio was involved — a play button so you can hear exactly what the character sounded like. Where the character used a tool, the call it made is shown inline.

This is the single most useful debugging page in the panel. "The character gave a wrong answer" is almost always answered by reading the transcript: either it misheard the question (visible in the transcript), or it had no document to answer from (fix on the Knowledge Base tab), or a guardrail blocked it (fix on Guardrails).

The search box matches speaker and character identifiers rather than names, so the usual route is to sort by time and open the conversation you are investigating.

Analytics — where the usage goes

Analytics page with interaction totals, token counts, daily charts and per-character breakdowns
Analytics over the last 7, 30, 90 or 365 days.
PanelWhat it tells you
InteractionsHow many turns were served in the period.
Tokens in / outThe volume of text going into and coming out of the model. This is the truest measure of load.
Speech secondsAudio generated and transcribed.
Daily interactionsYour traffic shape — when your characters are actually busy.
Daily tokensInput against output. A big input number usually means a lot of documents are being retrieved.
By characterWhich of your characters is carrying the traffic.
By modelWhich models are being used, and how heavily.
Analytics answers "how much"; Plan & Quota answers "how much is left". They count different things on purpose — remember that a turn on a flagship or real-time model consumes several times more of the allowance than the same turn on a small one.

Tool execution log

Tool Registry page listing defined tools with categories and endpoints, and the execution log button
Tool Registry. The Execution Log button opens the audit trail of everything the characters have called.

If your characters perform actions — sending mail, checking balances, calling your API — the execution log is the record of what was called, by which character, on whose behalf, and with what arguments. Each row links back to the conversation that triggered it, so you can read the exchange that led to the action.

Keep this open the first time you put a new tool live. Watching the first few real calls tells you very quickly whether the model is filling the parameters the way you intended, or whether your parameter descriptions need to be clearer.

A weekly routine that catches most problems

  1. Open Analytics on 7 days. Look for a spike or a collapse in daily interactions — both mean something changed.
  2. Check Plan & Quota. If a bar is heading for amber before the month is out, look at which character and model are responsible on the Analytics breakdowns.
  3. Read five random sessions. Not the complaints — random ones. It is the most reliable way to notice a character that has quietly started answering badly.
  4. Skim the execution log for failed actions.

Where to go next

Understand the allowance

Plan & quota →

Improve the answers

Knowledge & tools →

Panel overview

Panel tour →