Setting your answer policy and retrieval grounding
Configuring how your assistant answers knowledge questions and how strictly it grounds them, with every row showing its effective value and source.
Every value on this screen used to be a hardcoded environment variable, keyed by tenant ID, that nobody could see the effective state of. Now it's rows an operator reads and writes directly — and every row tells you not just what's set, but what's actually running.
1. Open City → Answer policy
You'll find three panels: Answering (Assistant), Grounding (Knowledge), and Answer cache — two different backend services and permissions, on one tab, because they're one conversation about how your city answers.
2. Read the effective value and layer before you touch anything
Every row shows the value your assistant is actually using right now, and whether it came from the platform default or your own tenant override. Leaving a field blank means "inherit the platform default" — it is not the same as setting a value that happens to match the default.
The confidence floor (contextReadFloor) is a safety threshold. Lowering it means your assistant answers on thinner evidence. Changing it, and changing the answer mode, both require a plain-language confirmation naming exactly what moves — never a bare "are you sure?"
3. Choose an answer mode deliberately
CoverageGated answers only when retrieval clears a coverage bar; ContextRead reads a fixed depth of context regardless. If you set a depth or a confidence floor while the mode is CoverageGated, the screen warns you that those fields are currently inert — the single most misleading state this configuration can be in is looking configured while behaving as if it isn't.
4. Tune grounding only if you've actually measured a regression
The reranker candidate pool, the distinctive-term threshold, and body-coherence checking are calibration switches, not everyday settings — which is why they require assistant.answer-policy.manage or knowledge.retrieval.manage, offered only to a tenant admin or platform admin, never a general editor role.
5. The answer cache is a separate lever: whether the same question gets asked twice
Below grounding, the Answer cache panel shows real entry, hit and anchor counts — not just settings — so you can tell an empty cache with the feature on from a full one that never actually gets hit.
Flushing the cache is deliberately outside the save form and behind its own confirmation. It strands every cached entry your city holds; that's not a setting, it's an action, and the screen treats it as one.