A concierge that already knows the wine list.

01 — The BriefAnswer the same fifty questions — beautifully.

Hearth & Vine had a problem the team called "the Tuesday list." Every Tuesday afternoon, a sommelier would clear ninety minutes to answer guest emails — wine pairings, menu allergies, late check-in requests, the same fifty questions, written fifty different ways.

The team didn't want a chatbot. They wanted the answers a senior sommelier would give at the door — informed, on-brand, never wrong about the cellar. They wanted it on the website by the holidays. And they wanted it to never invent a vintage.

02 — The SystemA small stack with one job.

I started by asking what not to build. No model fine-tuning. No bespoke vector database. No retrieval framework on top of a retrieval framework. I picked, in order:

• A single retrieval-augmented chat surface with a fixed system prompt, written by hand and reviewed by the head sommelier.
• A live corpus of the cellar (CSV, updated nightly), the menu (markdown, edited in the kitchen), and the booking calendar (read-only, scoped per property).
• An admin page that answers a single question for the team: what did the assistant get wrong this week?

The cellar corpus turned out to be the load-bearing piece. I standardised on a single line per bottle — vineyard, vintage, varietal, on-hand count, paired-dish notes — and let the retriever do the rest. The model never sees the whole cellar; it sees the eight or twelve lines that matter for the question being asked.

We stopped explaining the wine list and started asking guests where they wanted to spend the evening. The assistant did the rest.— Marin S., Head Sommelier, Hearth & Vine

03 — Refusal as a featureIt says I'm not sure when it isn't.

The single most contested decision was also the one the team thanked me for first: when the retriever doesn't surface a confident match, the assistant says so, and routes the guest to a sommelier with a one-tap button. No bluff, no improvisation, no soft-pedalling.

It costs the brand a percentage point of "self-served" interactions. It earns them the trust they were after in the first place.

04 — ReflectionWhat I'd do the same. What I wouldn't.

The same: keep the cellar as the only source of truth. Keep the prompt short and human. Keep the admin page boring on purpose — the team should never wonder what's underneath.

What I'd do differently: write the refusal language with the team on day one. The two weeks I spent rewriting it at launch were avoidable. Refusal is brand work; treat it that way.