Answers only from what you approved. Sources under every one.

Pages and documents are ingested into versioned articles with a status. On public channels only approved content is searched. Re-crawling a changed page bumps its version and drops it back to review, so superseded text stops answering the moment the source changed.

Your team can write the answer to a question once, as an answer card. When a question matches a card closely enough, the card's text is returned verbatim — no model rewrites it. A whole FAQ can be imported in one step.

The passages the answer was built from are shown under it as links to the page and section — one link per page, never five passages of the same article. The chat widget also lists which sources the assistant is allowed to search, so a resident knows where the answers come from before asking.

When nothing approved matches, the assistant abstains and offers a person — it never falls back to what a model happens to know. Each abstention is recorded as a gap in a review queue, so the missing page gets written rather than guessed at.

Each knowledge source carries a precedence. When two approved passages answer a question differently, the more authoritative one wins, the answer says another source differs, and a conflict item lands in the same review queue with both pages named.

Facts from a person's own record — an account balance, a parcel's assessment — appear in a public-channel answer only after that person has proven identity with what the record already holds. Staff see records in their own workspace; the public path never does without the check.

Rules written in plain English — "anything about a payment dispute goes to Billing", "never discuss an open court case" — are compiled into topic and signal matches that refuse, escalate or route. A policy can never add knowledge or unlock record data; it can only narrow what the assistant does.

A channel starts in shadow mode: every question is answered to a private log while staff keep answering, and reviewers grade the pairs. Your own question set runs as an accuracy evaluation — pass phrases, expected sources, expected abstentions — with a model judge as a second opinion on every answer, nightly if you want it. The bar is 95% by default; you set it.

"Was this helpful?" under every answer. A no asks why — wrong, out of date, didn't answer — and opens a review item carrying the question, the answer and the passages it used, so a wrong answer becomes a fix to a page, not a lost complaint.

Each source and article can name an owner and a validation cadence. Overdue or expired approved content is flagged to its owner; a page that has vanished from your website is flagged too, and the article keeps answering only until a person decides. Nothing silently goes stale.