Skip to main content

AI usage

Monitoring → AI usage shows how the AI assistant is used: who asks, how many tokens it costs, which tools it calls and which fail. It is for superusers, from Liberty Next → Operate → Monitoring.


At a glance​

MonitoringLiveServer logAI usageRetentionLast 30 days ▾Recorded: who, model, tokens, tools — never the question, the answer or a row. Kept 90 days.QUESTIONS412USERS18TOKENS IN9.6 M71 % from cacheTOKENS OUT184 kTOOL CALLS1 03812 failedAVERAGE TIME9.4 sper questionBY USERUSERQUESTIONSTOKENS INTOOL CALLSalice1483.4 M402bruno962.1 M233chloe611.5 M170BY TOOLTOOLCALLSFAILEDAVERAGE TIMEsql_query52271.2 sexplorer_query30130.9 ssearch_catalog11800.1 sLATEST QUESTIONSWHENUSERAPPLICATIONMODELTOKENS INTOOL CALLSTIMEOUTCOME18/11 10:42alicesalesclaude-sonnet-5-524 31038.1 sAnswered18/11 10:37brunosalesmcp—21.4 sAnswered18/11 10:31chloecrmclaude-sonnet-5-531 902514.6 sFailed18/11 10:20alicesalesclaude-sonnet-5-512 00413.0 sAbandonedAI usagea view of the Monitoring pageA click on a usernarrows the questions belowA row opensits tool calls, one by onePeriod7, 30 or 90 daysThe period in figuresquestions, tokens, tools, timeModel: mcpa call from an external client

What is recorded​

Each question to the assistant leaves one record: who asked, from which application's page, the model, the tokens (in, out, read from and written to the cache), the time it took and how it ended. Each tool call of the question is kept with its name, what it was called on (the connector and the query, the model, the keywords), whether it failed and why, its size and its time.

The question, the answer and the rows a tool returned are never recorded. A question the user abandons half-way is recorded with what had happened until then. A call from an external client is recorded the same way, with the model mcp — see Connect Claude (MCP).


The period in figures​

The period is Last 7 days, Last 30 days (default) or Last 90 days.

TileShows
QuestionsQuestions asked over the period.
UsersUsers who asked at least one.
Tokens inInput tokens, with the share read from the cache (71 % from cache): the part of a conversation sent again at each tool round is read from the cache, at a lower price.
Tokens outOutput tokens.
Tool callsTools called, with the number that failed.
Average time per questionFrom the question to the end of the answer.

Breakdowns​

TableColumns
By user, By day, By modelQuestions, Tokens in, From cache, Tokens out, Tool calls.
By toolCalls, Failed, Average time, Average size.

A click on a user narrows the questions below to theirs (Questions of …).


Latest questions​

One row per question: When, User, Application, Model, Tokens in, From cache, Tokens out, Tool calls, Time and Outcome — Answered, Failed or Abandoned. A row opens on its tool calls, each with its size in characters and its time.


How long it is kept​

Usage record kept (days; 0 = without limit), in Liberty Next → Administer → App → AI Assistant, is 90 by default. Older questions are deleted with the notifications' clean-up, and the store appears as AI usage on the Retention view, with its number of questions, Preview and Run now.


Tips​

  • A tool that fails often points at a missing permission or at a query whose description misleads the assistant: open the question's tool calls to read the reason.
  • A user with many questions and few tool calls is probably asking for what the assistant cannot reach: check the role's sql: and explorer: permissions.
  • The same figures are served by GET /admin/ai-usage for an external report.