Every call to an AI provider is recorded in a log: date, feature, provider and model, user, number of tokens (the providers' unit of account) and success or failure. The log keeps no content: neither prompts nor answers. The AI usage screen summarises it and lets you set caps.
It is reached through the AI usage button of the AI configuration screen, or from the Usage screen. It is reserved for Administrators.
Read the dashboard
Choose the Period (months): 1, 3, 6 or 12 calendar months, current month included. The screen shows:
- the totals: Tokens (input + output), Input tokens (the text sent), Output tokens (the text produced) and the number of Calls, including the failed ones;
- the trend By month;
- the breakdown By capability: chat, embeddings, document reading, image, with the number of calls, the tokens and, for the last two, the pages read or images produced;
- the breakdown By feature (the module and assistant behind the call, under a technical name), By provider and model and By user. Calls made by a background task, with no user, appear as System task.
Embedding tokens are estimated when the provider does not return them.
Export (CSV) downloads, for the chosen period, a file to open in a spreadsheet (semicolon separator, column headers in English): one row per month, feature, provider, model, user and capability, with the number of calls, failures and tokens, then the capability (capability) and the pages or images (units).
Set caps
Caps are set on the Usage screen (menu Administration › Instance and operations › Usage), in the Artificial intelligence section:
Monthly cap (tokens): for the whole instance;
Monthly cap per user (tokens): for each user;
Block chat at the cap: Yes refuses chat calls once a cap is reached, until the following month; No or (default) (which means No): the cap only raises an alert. Embeddings (semantic search) are never blocked.
Monthly cap (pages read) and Monthly cap (images produced): for document reading and for image. Once set, they always block when reached, whatever Block chat at the cap says. The cap is checked before each call: the call that crosses it (a long PDF, for example) runs to completion.
A cap specific to a feature is set in the Features tab of the AI configuration screen; it also blocks once reached. Blocking chat at the token cap also applies to document reading and image, which use tokens.
Leave a cap empty to set none. The alert threshold is 80 % of the monthly cap; it is changed with the AI:Usage:AlertPercent key on the Configuration screen.
When a global monthly cap is set, the AI usage screen shows a Monthly cap gauge: N / cap, then "Alert threshold crossed", "Cap reached (alert only)" or "Cap reached: AI chat calls are blocked". In the assistants, a user blocked by a cap sees the message "This month's AI usage cap has been reached. It is lifted at the start of next month, or earlier if an administrator raises it.": a message distinct from the one shown when AI is not configured.
With the AI provided by the publisher, document reading and image are deducted from the same token credit as chat, and refused once the credit is used up. When the service does not return the number of tokens of a call, an estimate is deducted: 1,500 tokens per page read, 4,200 per image produced.
Log retention
The log is kept for 365 days, then purged every night by the AiUsagePurge scheduled task. The duration is changed on the Usage screen or with the AI:Usage:RetentionDays key.
When AI is disabled
If chat is set to None while an embedding provider is still chosen, the screen says so ("The AI chat is set to "None": no chat call is made. Semantic search (embeddings) stays on: its calls are still logged here. Any history remains available."): semantic search keeps calling the AI and being logged. If no provider is chosen at all, neither chat nor embeddings, the screen shows "AI is disabled: no calls are made and nothing is logged. Any past history remains available." and nothing more is added to the log. In both cases, the past calls still kept remain displayed.
See also
- "Configure AI": choose and set up the providers.
- "Monitoring usage": the Usage screen, its caps and alerts.