Grounded and cited — or it refuses

Aria answers your guests from your own content, and shows where the answer came from. When it is not in there, she says so instead of guessing. These are the numbers she has to hit before any change reaches your hotel.

Last verified: Aug 17, 2026 · Engine build: fv-grounding-20260816b

What the last gate run measured

Every figure below comes from that one run. Nothing on this page is an estimate.

100.0%

Search hit rate (top 8)

How often the passage holding the real answer is among the top 8 passages Aria reads before she replies.

Deploy blocked below: 90.0%

100.0%

Chat answer pass rate

Share of 29 scripted questions where Aria's reply matched the true answer.

Deploy blocked below: 95.0%

100.0%

Refusal accuracy

Share of 18 deliberately made-up questions Aria refused instead of inventing an answer.

Deploy blocked below: 90.0%

Enforced

Hotel-to-hotel isolation

Your content answers your guests only. A question aimed at another property's knowledge comes back with nothing at all.

Passages returned from a property that is not yours: 0

The rest of the scorecard

Rank of the right source

0.986

How near the top the correct passage sits, averaged over every question. 1.000 means it was first every single time.

Wrongly refused

0.0%

Questions Aria refused even though the answer was in the content. Lower is better.

Scored on 90 questions written against real hotel content: 72 with a correct answer to find, and 18 invented questions that must be refused.

Try to trick her

The fastest way to judge a hotel chatbot is to ask it something it cannot know. Ask Aria about a spa nobody mentioned, a rate nobody published, or a policy that does not exist — and watch her refuse instead of inventing one.

Ask Aria something she shouldn't know

How these numbers are produced

Aria is not asked to be clever. She retrieves passages from your own content and answers from those passages only, with the source shown next to the reply. If nothing relevant comes back, she refuses and offers a human instead.

Every change to the engine, the prompts or the knowledge base is run against a fixed set of questions before it ships. Part of that set has a correct answer to find; the rest is invented on purpose and must be refused. If any gate on this page is missed, the deploy stops and the change never reaches your hotel.

The same run also checks that one property's knowledge can never answer another property's guests, and that a cached reply still matches a freshly computed one. The date above is the date of the run these figures come from.