Guides

Every procedure in this documentation lives on the page for the screen it belongs to. This page puts those procedures in the order you are likely to need them, so you can work through a task without knowing which screen owns it.

1. Get AskRAI answering questions

Nothing works until a model is connected and there is content to answer from.

1. Add a model provider — AskRAI runs on your own provider accounts, so this comes first. Test the model before moving on.
2. Route that model to each pipeline stage — summarization, query refinement, guardrail judging and confidence scoring can each use a different model. Nothing takes effect until you publish.
3. Create a knowledge pack and upload files to it — the pack's content type is fixed at creation, so choose it deliberately.
4. Write a Q&A pair by hand where an uploaded document does not cover the question. A new pair starts as a draft and answers nothing until approved.
5. Ask a question in the Sandbox — it runs the full pipeline without reaching an end user, and shows you which knowledge items the question retrieved.

2. Decide what it is allowed to say

Governance is attached to a subject — a group, an agent, or a policy — never to the tenant at large.

1. Understand what a guardrail is before writing one: a kind, a stage and an action, decided independently.
2. Author a guardrail — this is the only surface carrying the four fields that decide whether it does anything.
3. Attach it to a group, an agent, or a policy — an unattached guardrail screens nothing, for anyone.
4. Set an egress allowlist if your agents call tools, to bound which hosts they may reach.
5. Tune the confidence bands, then add an escalation rule so a weak answer raises a ticket.
6. Review how each guardrail is performing once real traffic has run through it, and mark false positives from the conversation logs.

3. Give people and programs access

1. Create the groups your organization actually uses — groups are what carry guardrails and knowledge packs to a person.
2. Provision users from your identity provider over SCIM 2.0 or Google Workspace, rather than maintaining them by hand.
3. Grant Platform Admin to the small group who configure the tenant. Everyone else is a Member, which needs no assignment.
4. Issue a credential for the API, CLI or MCP server. The secret is shown once.

4. Run a team of agents

1. Create an organization, then a project inside it.
2. Allocate a budget and an idle check-in — the first bounds what the project may spend in total, the second bounds how fast an idle team keeps spending.
3. Add people and agents to the team, and set each agent's reporting line so the org chart and manager oversight are right.
4. Set each agent's cadence and autonomy — how often it acts, and how much it may do without you.
5. Start the team and watch it work — the live graph, the message thread, the task board and the approval queue.

5. Watch it in production

1. The dashboard for volume, response time, success rate and the mix of human, application and agent traffic.
2. Conversation logs to read individual exchanges, inspect what each check decided, and mark false positives.
3. Tickets for the escalations your rules raised, with the exchange that triggered each one attached.
4. Governance & audit for what the audit record keeps — and the two things it deliberately does not.

Related pages