Designing Usher to say "I don't know"

Website chatbots fail in a predictable way: asked something outside their knowledge, they improvise. For a support widget that's embarrassing; for a product answering pricing or compliance questions, it's a liability. Usher is being built with refusal as a first-class outcome, not an error state.

Evidence-gated answers

Usher only answers when retrieval produces passages that actually support the answer, and every response carries its citations — the exact page sections it drew from. If the evidence is thin, contradictory, or absent, the honest move is a handoff: 'I don't have a reliable answer; here's a human who does,' with the conversation context attached so the visitor never repeats themselves.

• Retrieval confidence thresholds gate generation — no passage, no answer
• Citations are rendered in the widget, so visitors can verify instead of trust
• Handoff routes carry full conversation context to email or a support desk (Relay, eventually)
• Every conversation is auditable by the site owner — it's their data, on their infrastructure

Usher pairs with Vyom for model routing, which means the same evidence-gating works across whichever models you've approved — and conversations never leave infrastructure you control, let alone train someone else's product.