We're building an answer engine for hotel rooms

The questions people actually ask about hotel rooms don't fit in a search form. "Vegas in March, six of us, we need a door between us." "Somewhere warm with two bedrooms that isn't Orlando." "A suite where the kids' bathroom isn't through our bedroom." There's no field for any of that, so people filter by city and price and then read room descriptions one at a time.

So we built Ask Suiteness: describe the trip in your own words and get real rooms back. This post covers how it works, why we built it this way, and what it can't do yet.

Why not just add a chatbot

A general-purpose chatbot answers hotel questions from what it absorbed in training. It writes confidently about suites that were renovated three years ago, rooms that don't exist, and prices it has no way of knowing. For most questions that's an inconvenience. When the question is "will there be a door between our rooms," a confident wrong answer ruins a holiday.

So we started from one rule: the model writes the sentence, and the database supplies the rooms.

How it works

When you ask a question, a language model reads it and decides what to look up. It can't answer from memory. It has three tools, and all three query our own catalogue:

• Search suites, by destination, hotel, party size, bedrooms, bathrooms, floor area, nightly budget, and whether the room is a guaranteed connecting pair.
• Get one suite in full, including the beds room by room, which is what answers "who sleeps where."
• List destinations, to turn "somewhere warm" into cities we actually have rooms in.

The model reads the results and writes a reply. Everything else on the page, including each suite's photo, hotel, bedroom and bathroom counts, size, nightly figure, link and the line explaining why it matched, is built directly from the rows the database returned. The model never writes those parts.

That split is the whole design. It means the page can't show a room that doesn't exist, quote a price nobody stored, or link to a page that isn't there, whatever the model says next to it. The prose can be clumsy. The rooms can't be invented.

The same answers in other assistants

The same three tools are available as a public MCP server at https://www.suiteness.com/mcp, so Claude, ChatGPT and other assistants can search the Suiteness catalogue directly instead of guessing. It's read-only and needs no account, because it serves the same public catalogue the website does. If you use Claude, add it as a custom connector and ask about your trip there.

What it still gets wrong

Being clear about this matters more than the feature list.

• Prices aren't quotes. The nightly figure is the lowest recent rate we've seen for that suite, not the price for your dates. Ask Suiteness says so, and every answer links through to search, where the live rate, taxes and fees are confirmed.
• It doesn't check availability. A suite that matches your description may be sold out on your dates. That's the next thing we want to connect.
• It only knows what we've recorded. It can filter on bedrooms, bathrooms, floor area, occupancy, price and connecting doors, because those are in the catalogue. "Walkable to the beach" and "step-free shower" aren't recorded, so it says so and points you to our concierge team instead of guessing.
• It's only as accurate as our data. If a suite's bathroom count is wrong in our catalogue, the answer will be wrong too. When guests catch those mistakes, we fix the data, which fixes the answers.
• Long questions take a couple of turns. To keep answers fast and costs bounded, it does a limited amount of searching per reply. A trip with five constraints may need a follow-up message to narrow things down.
• It forgets. Conversations stay in your browser tab and aren't stored. That's deliberate, but it means closing the tab starts you over.

Why we think it's worth it

People who book suites and connecting rooms are usually planning for a family or a group, and they have specific questions. A search form makes them translate those questions into filters. An answer engine built on real inventory lets them just ask. If you have a trip coming up, try it and tell us where it falls short.