⛨ Grounded + gated
Retrieval first, model second. Weak retrieval → honest "I don't have that" + human escalation. It never guesses.
❞ Cited
Every answer carries the article it came from — your team can audit any reply.
⚡ Streaming
Tokens stream in live (SSE). First words in under a second on a warm model.
☂ Never dies
If the LLM is unavailable, it degrades to verbatim quoted excerpts from the source — labeled, still cited.
Poke at it
Green questions are in the knowledge base. The red one isn't — that's the point: watch it refuse and offer a human instead of inventing an answer.
⚠
Full disclosure: "Solstice CRM" is a fictional product — the 12-article knowledge base was written for
this demo. The escalation form acknowledges but doesn't send real email. Generation runs on a free-tier model cascade;
when quotas run dry the widget degrades to labeled verbatim excerpts instead of failing. In a client install, the same
engine sits on your real help center with pgvector/embedding retrieval and your helpdesk hooked to escalations.