Roles and permissions
How system administrators manage built-in and global roles, edit permission titles, and set platform approval rules under /admin/roles.
Knowledge
How system administrators manage platform-wide knowledge collections under /admin/knowledge, and how ingestion and hybrid retrieval work for AI chat.
Roles and permissions
How system administrators manage built-in and global roles, edit permission titles, and set platform approval rules under /admin/roles.
Knowledge
How system administrators manage platform-wide knowledge collections under /admin/knowledge, and how ingestion and hybrid retrieval work for AI chat.
How system administrators enable AI Gateway models, pick default models per use case, edit per-function instructions, and set the agent policy and runtime defaults under /admin/ai.
AI settings is the platform-level AI configuration at /admin/ai, under Admin > Operations > AI. It has three tabs:
| Tab | Route | Purpose |
|---|---|---|
| Models | /admin/ai | Enable models and choose the default model per use case. |
| Instructions | /admin/ai/instructions | Edit system instructions and generation settings per AI function. |
| Policy | /admin/ai/policy | Set the agent execution policy and runtime defaults. |
These settings apply to every organization. Organization-level AI preferences and user-facing chat are described in Chat.
provider/model. Only enabled models can be used.chat.ask, chat.agent or contextual.customer.summarize. Each function has English system instructions and generation settings.The Models tab lists the AI Gateway catalog grouped by provider. Each provider row shows how many of its models are enabled, for example "4 of 84 enabled". Expand a provider to see its models with context window, release date, capability tags (shown in the Permissions column), and input and output cost.
Toggle Enable model on a row to enable or disable it. Use Search providers or models… and Filters (Provider, Permissions, Status) to narrow the list.
The Default models card has one selector per use case. Pick an enabled model for each; you must enable at least one model before you can set a default.
The Chat default also powers the public docs assistant on the documentation site. The docs app reads the enabled Chat default and refreshes it every five minutes. When no enabled Chat default exists, the docs assistant falls back to google/gemini-3.7-flash.
Embedding defaults must produce 1536-dimensional vectors, because knowledge chunks are stored in that size. See Knowledge.
The Instructions tab lists every AI function with its name, key, category, status and last update. Categories include Chat, Contextual, Suggestions and Recognition. Filter by category or status, or search by name.
Open a function at /admin/ai/instructions/{functionKey}, for example /admin/ai/instructions/chat.ask. Depending on the function, you can edit:
| Field | Purpose |
|---|---|
| System instructions | The English system prompt for this function. |
| Input template | The constructed input text sent with each request. |
| Max steps | The maximum number of tool-calling steps. |
| Max output tokens | The cap on generated tokens. |
| Timeout (ms) | How long a call may run. |
| Temperature | Optional sampling temperature. Empty uses the provider default. |
Select Save to store an override, or Reset to default to remove it. Instructions stay in English. The output language and date and time formatting are applied at runtime from the signed-in user's preferences.
The Agent policy card on the Policy tab controls what Agent chats may do:
Ordinary create, update and delete tools run without approval unless they are high risk or always gated.
The Runtime defaults card sets platform defaults that users can still override per turn:
| Setting | Default |
|---|---|
| Default composer mode (Ask, Plan or Agent) | Agent |
| Default reasoning | high |
| Default speed (Normal or Fast) | Normal |
| Default context window | 200,000 tokens |
| Suggestion count | 3 |
| Suggestion min words and max words | 8 and 22 |
| Knowledge retrieve limit and preload limit | 8 and 6 |
| Memory always-on limit and semantic limit | 4 and 6 |
| Memory min confidence | 0.85 |
The agent policy and runtime defaults are stored as the ai_agent_policy and ai_runtime_defaults keys in platform settings.
| Key | Effect |
|---|---|
system.admin.access | Opens the AI tabs and is checked by the model, default and instruction actions. |
system.ai-models.manage | Required by the database to write the agent policy and runtime defaults. |
Last updated on
Roles and permissions
How system administrators manage built-in and global roles, edit permission titles, and set platform approval rules under /admin/roles.
Knowledge
How system administrators manage platform-wide knowledge collections under /admin/knowledge, and how ingestion and hybrid retrieval work for AI chat.