AI Policy and Limits
Settings → AI → Policy holds the decisions usually made once, with security or legal input, and then left alone.
Opening and saving this tab needs the AI settings permission — see AI Roles and Permissions.
Values apply to the whole installation. 0 means no limit, everywhere it appears.
Privacy & safety
Message payloads and password fields are always masked before the AI sees them. This cannot be turned off.
| Setting | Default | What it does |
|---|---|---|
| Tell the AI when input is masked | Off | Adds a note to the model's instructions so it treats a ██████ as redacted on purpose. Doesn't change what's sent |
| Semantic document search | Off | Search documents by meaning, not just words. Also sends document text to an embedding provider |
Document search today is full-text. Leave this off unless your vendor has advised otherwise.
Retention
| Setting | Default | What it does |
|---|---|---|
| Conversations (days) | 90 | Conversations and their attachments are deleted after this many days |
| Audit records (days) | 400 | How long a record of what the AI did is kept. Usually longer than conversations |
Budget
| Setting | Default | What it does |
|---|---|---|
| Monthly tokens per user | 0 (no limit) | Caps how many tokens one user may spend in a period. Warns at 80%, refuses at the limit |
The period is a calendar month, measured in UTC.
When a user reaches the limit, the run is refused with a message saying the budget resets at the start of the next period. Runs already in progress are unaffected.
Project documents
| Setting | Default | What it does |
|---|---|---|
| Documents per project | 25 | How many documents one project may hold |
| Max document size (MB) | 25 | Largest single document that may be uploaded |
Documents are never removed by retention, so these two are what bound database growth. Re-uploading content the project already holds does not count against the document limit.
Chat attachments
| Setting | Default | What it does |
|---|---|---|
| Max attachment size (MB) | 10 | Largest single file a user may attach to a message |
| Attachments per conversation | 10 | How many files one conversation may carry |
| Total attachment size per conversation (MB) | 40 | Combined size of a conversation's attachments |
| Conversation size kept (MB) | 1 | How much of a conversation is kept |
Past this size the oldest messages are dropped so the recent conversation survives. At the default, 1 MB is roughly a quarter-million words.
Reference budget
A reference is material added to a request beyond the conversation itself — a Skill's instructions, a retrieved document passage, or a file named in the request.
| Setting | Default | What it does |
|---|---|---|
| References per request | 5 | How many references one request may carry. Extras are left out and the chat says which |
| Reference share of context (%) | 40 | The most of the model's context window references may fill |
There is no size or count limit on Skills themselves. What actually bounds how much Skill text reaches a request is the reference budget above. See Organization Skills.
Always in effect
Four things hold regardless of how this tab is set, and the UI states them:
- Payloads and passwords are masked before the AI sees them.
- Generated code is validated before it runs.
- Agent telemetry never leaves this site.
- A model reaches the internet only if you turn on web search for it (Settings → AI → Setup).
What administrators can review
Settings → AI → Usage shows token usage per user for a date range — Today, Last 7 days, This month, All time, or a custom range. Each row expands to a per-model breakdown. A user sees only their own usage unless they hold AI settings.
A Reset on that tab, which also needs AI settings, moves the baseline so the tab shows nothing recorded before that moment. It does not delete records and it does not affect budgets.
Separately, AI system events appear on the Logs page as SYSTEM_EVENT records with event types beginning AI_:
| Recorded | What it tells you |
|---|---|
| Run usage | Tokens a run consumed |
| Quota warning | A user passed 80% of their monthly budget |
| Quota exceeded | A run was refused because the budget was spent |
| A relaxed policy value | A limit on this page was loosened |
| A blocked run | A run the platform refused |
Loosening a policy value is recorded with who changed it and the before/after. Tightening a value is not.
If it didn't work
| Symptom | Cause | Fix |
|---|---|---|
| An upload is refused with a size or count message | The file or the project is past one of the document or attachment caps | Raise the cap the message names, or split the upload |
| The chat says a reference was left out | The request needed more references than the budget allows, or they would fill too much of the context window | Raise References per request or Reference share of context (%), or attach less |
| A run is refused with a message about the budget resetting | The user has spent their Monthly tokens per user allowance | Raise the allowance, set it to 0 for no limit, or wait for the next calendar month |
| The Policy tab is read-only or missing | The account lacks AI settings | Grant the permission to the role |