AI assistant
Overview
The AI assistant is the internal AI workspace. You ask a question or describe a need, and the answer is built from this system's data — not from general world knowledge.
That is the essential difference from any chat tool: the assistant sees only what you are allowed to see anyway. Your permissions continue to apply unchanged; the AI is given no separate, wider view of the data.
Core tasks
Hold a conversation. The list of conversations sits on the left, grouped into Today, Yesterday, This week and Older, and searchable. New conversation starts a fresh thread.
Ask questions. Enter sends, Shift + Enter inserts a new line. On an empty thread you are offered suggestions that show what the assistant is meant for — summarising open work, drafting a reply to a customer, explaining how to do something.
Check the sources. Answers list the sources they used. This is the single most important habit when working with the assistant: an answer without a traceable source is an assertion, not information — the citation is what makes it checkable.
Choose the model. The model picker lets you pin a specific language model; the default, Automatic choice, uses the route configured for this use case (Settings → AI Settings, tab Assignment per task), otherwise the system default model. It does not pick a model by price or context length.
Use an agent or building block. The input bar lets you pick an agent (a preconfigured assistant with its own instruction and allowed tools) and a skill/prompt (a text building block inserted into the input field).
Maintain agents and building blocks. The menu behind the three dots carries both: New agent creates one in a short dialog, Manage agents & prompts opens the full maintenance screen in the AI settings — that is where entries are changed, shared, published and deleted. That tab needs no special permission: agents and building blocks belong to the user who creates them.
Manage conversations. Threads can be renamed and deleted.
Fields in detail
The What it does column answers what changes in the system — not what the field is called. The grey name behind it is the API field: the same thing runs under that name through automation, import and AI tools.
Writing a message
| Field | Required | Values / format | What it does |
|---|---|---|---|
Message content | yes | text, max. 20,000 characters | Your question or instruction. The assistant reads only what you may see anyway — the permission check sits in the interface, not in how the request is phrased. |
Conversation threadUlid | yes | existing conversation | Which thread the message belongs to. On the first send the interface creates the thread itself. |
Conversation
| Field | Required | Values / format | What it does |
|---|---|---|---|
Title title | no | text, max. 255 characters | What the thread is called in the list. It comes from the first question and can be renamed — the automatic title rarely fits for long. |
Space spaceUlid | no | existing space | Which working space the thread sits in. The list on the left then shows only that space's threads. |
Brand brandId | no | one of your brands | Which brand the thread refers to. |
Model modelRoute | no | model from the catalogue, max. 100 characters, empty = automatic choice | Which language model answers. Empty leaves the decision to the system; the choice applies from the thread's next message on. |
Space
| Field | Required | Values / format | What it does |
|---|---|---|---|
Name name | yes | text, max. 150 characters | What the space is called in the switcher above the thread list. |
Visibility visibility | no | private, team, brand (default: private) | Who sees the space. Created through the interface it is always private; sharing runs through the API. |
Brand brandId | no | one of your brands | Binds the space to a brand. |
Icon icon | no | text, max. 50 characters | The mark of the space. |
Colour color | no | text, max. 20 characters | The accent colour of the space. |
Reference type entityClass | no | text, max. 255 characters | What kind of record the space hangs on — a project or a customer, say. |
Reference record entityId | no | number | Which record exactly. Together with the reference type this turns the space into the AI surface for that one case. |
Agent
An agent is a preconfigured assistant: its own instruction, its own tools, its own rule for what it may do without asking. New agent sets name, instruction, tools, sub-agents and autonomy; everything else lives in the AI settings.
| Field | Required | Values / format | What it does |
|---|---|---|---|
Name name | yes | text, max. 150 characters | What the agent is called in the composer's picker. |
Description description | no | text | What it is meant for. Pure explanation for humans. |
Instruction instructionInline | no | text | The role in your own words: what the agent should do, what it watches for, what it leaves alone. |
Instruction preset instructionUseCase | no | key of a use case, max. 100 characters | A centrally maintained instruction instead of your own text — so the same role does not exist in twenty versions. |
Tools allowedTools | no | list of released tools | What the agent may do. Without a tool it is purely advisory. A tool never grants more than your account may. |
Sub-agents allowedAgents | no | up to 8 agents | Whom it may hand part of a job to. Only what you can reach yourself is accepted — an agent never gets a wider reach than its user. |
Autonomy autonomy | no | advisory (advise only), confirm (write only after asking), whitelisted (write within the tool list without asking) | How far it may act without you. whitelisted is an administrative decision, not a personal setting. |
Knowledge policy brainPolicyUseCase | no | key of a use case, max. 100 characters | Which slice of knowledge the agent answers from. |
Model modelRoute | no | model from the catalogue, max. 100 characters | Which language model it uses. Empty means automatic choice. |
Status status | no | draft, published, archived | Whether it is only a draft, usable, or taken out of service. |
Visibility visibility | no | private, team, brand | Who may use it. It is created as yours; sharing is a separate step. |
Brand brandId | no | one of your brands | Binds the agent to a brand. |
Own model endpoint
Manage right only: an additional provider or a model in your own network.
| Field | Required | Values / format | What it does |
|---|---|---|---|
Name name | yes | text, max. 100 characters | How the endpoint appears in the model list. |
Kind type | no | openai_compatible, ollama, openrouter | Which protocol is spoken. |
Address baseUrl | yes | address (URL), max. 255 characters | Where the requests go. The address is checked: targets inside your own network are blocked unless explicitly released. |
Credential credential | no | text | The provider's key. It is stored encrypted and never returned — only whether one exists is shown. |
Models models | no | list of model names | Which models this endpoint offers. |
Use cases allowedUseCases | no | list of use cases | What the endpoint may be used for. Empty means all. |
Enabled enabled | no | yes/no | Whether the endpoint contributes models. |
Allow target in own network allowPrivateHost | no | yes/no | The explicit release for a local model. Valid for this endpoint only; cloud metadata and reserved ranges stay blocked even so. |
Brand brandId | no | one of your brands | Binds the endpoint to a brand. |
What the assistant may do
| Element | What it means |
|---|---|
| Knowledge source | Answers rest on the system's knowledge and on the help — with the source named. Without a source it says so instead of guessing. |
| Actions | It can act, but only through explicitly released tools and in your name: what you may not do, neither can it. Every action appears in the audit log, marked as AI. |
| Text snippet | A skill/prompt is inserted into the input field — it is a text suggestion, not a separate instruction to the model. It is maintained in the AI settings. |
| Cost | Every run is recorded with its consumption. That keeps the cost of use visible instead of surprising you at month's end. |
Settings & permissions
- Two feature flags. The assistant requires
ai.gatewayandai.workspace. If either is
missing, the page shows a notice instead of an input line — switched off means switched off here, not "quietly running in the background".
- Providers and models are configured centrally. The available models come from a catalogue that
is refreshed regularly; they are not hard-coded.
- Your permissions apply. The assistant can see nothing and do nothing your account is not
entitled to.
- Agents and building blocks belong to the user. Everybody creates their own; the visibility
Team or Brand shares them with others. What an agent may do without asking, by contrast, is decided by administration.
- Actions are traceable. Whatever the assistant changes appears in the
audit log with the AI as the actor — kept apart from human actions.
FAQ & troubleshooting
The page says the AI workspace is not enabled. Then at least one of the two feature flags ai.gateway and ai.workspace is missing. Your administrator sets those.
The assistant does not know data that clearly exists. Either you lack the permission for it — in which case that is correct behaviour — or that area is not set up as a knowledge source.
The answer cites no source. Then it is not grounded in a retrieved record. Treat it with corresponding caution and verify via Search.
A conversation has disappeared. Threads can be deleted; deletion is final and is confirmed beforehand.
Where do I change an agent I created? Via Manage agents & prompts in the menu behind the three dots. The New agent dialog only creates.