How the AI answers guest questions

Every building has its own knowledge base, the house rules, amenities, check-in details, and documents you provide during onboarding. When a guest texts a question, Rezi doesn’t make something up. It searches your knowledge base for the most relevant facts and answers using only that information.

This architecture has a name in the AI world, retrieval-augmented generation, but the idea is simple enough to explain over coffee: instead of letting a language model answer from its general training (which knows nothing true about your building), Rezi first retrieves the specific facts you wrote down, then has the model compose a reply using only those facts. The model supplies the fluent, natural phrasing; your knowledge base supplies every piece of substance.

Why build it this way? Because the failure mode of a free-wheeling AI, confidently inventing a door code or a pet policy, is unacceptable when real guests act on the answers. Grounding the AI in your facts trades a little flexibility for a lot of trustworthiness: Rezi would rather say “let me check with the manager” than guess.

Step by step

1. 1

   You provide the knowledge

   During onboarding you add facts and upload documents. Rezi converts each into a vector embedding and stores it.

2. 2

   A guest asks a question

   The incoming text is also converted to an embedding so it can be compared by meaning, not just keywords.

3. 3

   Rezi finds the best matches

   A similarity search pulls the handful of knowledge entries most relevant to the question.

4. 4

   The AI writes the answer

   It composes a natural reply grounded in those entries, in the guest’s language and the assistant’s consistent built-in voice.

What “matching by meaning” buys you

The embedding step is what makes the search smart rather than literal. A guest who texts “my internet is down” will match your entry about the WiFi network even though the words don’t overlap, because the system compares meaning, not spelling. The same goes for typos, slang, and other languages: “se rompió el aire” finds your air-conditioning entry just as reliably as plain English would. You write each fact once, naturally, and it’s reachable from every direction a guest might approach it.

It also means you don’t need to anticipate every phrasing. Operators sometimes try to write entries like a list of FAQ permutations; that’s unnecessary. One clear statement, “Quiet hours are 10pm to 8am, Sunday through Thursday”, handles “when do quiet hours start,” “can I play music at midnight,” and “neighbors are loud, what are the rules” equally well.

One brain, every channel

Text messages and voice calls run on the same retrieval system over the same knowledge base, so a caller and a texter get the same answer to the same question. Update a fact once and both channels pick it up immediately, there’s no separate “phone script” to maintain. The assistant’s built-in identity gives every reply one consistent voice, while the knowledge base controls what it says.

Why answers improve over time

The better and more complete your knowledge base, the better the answers. When you notice a question Rezi couldn’t answer, add that fact and it will handle it the next time. See improving answer accuracy.

In practice this creates a satisfying loop: the questions that escalate in week one are precisely the gaps in your knowledge base, and each one you fill is a conversation you never see again. Most operators find the system reaches a steady state within a few weeks, where the only escalations left are the ones that genuinely should reach a human.