Sessions & analytics
Every conversation, in full — plus where your usage is actually going.
Once characters are live, most of your time in the studio is spent here. Both pages are under Insights in the top bar: Sessions tells you what was said and why an answer went wrong; Analytics tells you which character is busy and which model uses the most. The audit trail of tool calls is under Knowledge › Tool Registry › Execution Log.
Sessions — the transcripts
Insights › Sessions lists every conversation across all your characters, newest first ("Find conversations by character, visitor or session ID."). A session can come from any place your characters talk: a website embed or widget, a kiosk, the Unreal plugin, a phone call, WhatsApp, or your own Test conversation in the editor.
- Characters on the side narrows the list to one character (All characters shows every one); Find a character… searches that list.
- The search box (Search character, visitor or session ID…) matches a character's name or ID, the visitor's ID and the session ID.
- Two filters: status (All statuses, Open, Closed) and activity (With messages, No messages). Open means the conversation is still running. Clear filters resets everything.
- Each row shows the character, the visitor, its status and the number of turns. Load more sessions fetches older ones; Refresh reloads the list.
Opening a session shows the whole exchange. The header has the speaker (visitor) ID, Started, Duration, Turns and Cost. Under Messages each message is marked USER, ASSISTANT, SYSTEM or TOOL, with its token counts and, where audio was stored, a player so you can hear exactly what the character said. A message that called tools has an expandable Tool calls list. Device Info shows what the client reported, for example the channel of a WhatsApp conversation. Delete removes the session after a confirmation.
Phone calls also have their own Call history with View transcript on Channels › Phone, and WhatsApp chats are answered live in the Channels › WhatsApp inbox.
Analytics — where the usage goes
Insights › Analytics ("Usage statistics and cost breakdown.") covers the Date range you choose: Last 7 days, Last 30 days, Last 90 days or Last year.
| Panel | What it tells you |
|---|---|
| Interactions | How many usage events were recorded in the period — mostly conversation turns, plus runs of Generate with AI: a plain count, not quota units. |
| Tokens In / Tokens Out | The volume of text going into and coming out of the models. A large input number usually means a lot of knowledge is being retrieved. |
| TTS+STT (sec) | Seconds of speech generated and transcribed. |
| Estimated cost | An estimate of the underlying model cost, for comparing characters and models. It is not your invoice and not your plan price. |
| Daily Interactions | Your traffic shape — when your characters are actually busy. |
| Daily Token Consumption | Input against output per day. |
| Interactions by Character | Which characters carry the traffic (the top ones). |
| Tokens by Model | How much each language model consumed. |
| Event Type Distribution | The kinds of usage events recorded. |
| Character Cost Ranking | Which characters cost the most. |
Tool execution log
If your characters perform actions — sending mail, checking balances, calling your API — open Knowledge › Tool Registry and press Execution Log. The Tool Execution Log ("Audit trail of every tool the LLM has called.") shows Total calls, Successful, Failed and Distinct tools, then every call, newest first.
- Filters: a search box (tool name, character, speaker, arguments), Filter by tool and a Date range.
- Each call shows its status (ok or error), the Arguments the model sent and the Result or Error.
- View session opens the conversation that triggered the call, so you can read the exchange that led to it.
A weekly routine that catches most problems
- Open Analytics on the last 7 days. Look for a spike or a collapse in daily interactions — both mean something changed.
- Check the plan menu or Plan & Quota. If a meter is heading for amber before the month is out, find the character and model responsible in the Analytics breakdowns.
- 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.
- Skim the execution log for failed calls.