Skip to content

How the AI uses your food & beverage knowledge

When a customer asks the menu AI “what should I drink with the carbonara?” — or when the cashier-side suggestion AI tells a server what to upsell — the AI is not searching the internet. It is reading three things you control. This page explains what those three things are, and where your input matters.

This page is for owners, chefs, and managers who want the AI to sound like their restaurant, not a generic assistant.

What it does

A good recommendation needs three kinds of knowledge:

  1. What’s available right now — the current menu, with tonight’s prices, what’s hidden, what’s out of stock.
  2. What things are made of — that burrata is dairy, that Sangiovese is a grape commonly used in Chianti, that this dish is from the south.
  3. How you would describe it yourself — the story of the dish, where the flour comes from, what the chef would pour with it, the way your team talks about food.

The first one is automatic — the AI reads it from the menu you maintain in the admin app. The second is also automatic — every time you edit a menu item, the system runs an extraction in the background that catalogues the dish (ingredients, allergens, region, technique). The third one is the part you write, and this folder is where it lives.

How to use it

Where to write your knowledge

There are two homes for your narrative, depending on whether the note is about one specific dish or your restaurant as a whole. Plus a shared library you don’t have to maintain.

  • Shared knowledge (every restaurant) — wine-pairing principles, what “al dente” means, what gluten is, how to describe a dish well. Lives in technical files maintained centrally. You don’t edit these.
  • Per-dish private notes — in the dish editor, alongside the public description, there is an Internal notes field. The AI reads it, the public menu never does. Use it for things the chef wants the waiter or AI to know about this specific dish but that don’t belong on the customer-facing menu (cross-contamination flags, upsell rationale, supplier caveats, internal vocabulary).
  • Your venue context — in Settings → AI context (planned), one markdown page. The chef’s mother is from Avellino. Your house style says “guanciale, not pancetta.” Your philosophy. Sourcing principles (why you picked your suppliers, not who — that’s already in your inventory). Anything that applies to your restaurant, not to a single dish.

What to write first — venue context

About 500–1500 words total. Use markdown ## headings to organise. Suggested sections:

  1. Philosophy. What your restaurant stands for, in the chef’s own words. Not marketing speak.
  2. Chef bio. Two or three sentences. Where they trained, what they cook, what they care about.
  3. Sourcing principles. The why of your suppliers. Don’t list who you buy from (that’s already in inventory) — explain why you chose them, what makes them different.
  4. House style. How your team describes food. The Italian terms you keep, the ones you translate. The level of detail you offer unprompted.
  5. Signature dish stories. One paragraph each for the 2–4 dishes you would tell a guest about unprompted.

What to write first — per-dish internal notes

Not every dish needs internal notes. Most dishes you’ll only fill in the public description that customers read. The 10–20% of dishes that do need internal notes are the ones with hidden context:

  • Allergen / cross-contamination flags — “shared fryer with breaded items”
  • Supplier caveats — “primary fornitore is Caseificio Stagni, secondary is X (slightly milder)”
  • Service notes — “two-person portion when ordered at table 14”
  • Upsell logic — “always pair with the Falanghina, not the Frascati”
  • Seasonal variation — “in winter we substitute the basil with parsley”

What to keep updating

  • Seasonal changes. Switch a sourcing line in AI context when your tomato supplier rotates; update per-dish notes when ingredients change.
  • New signature dishes. Add a paragraph in AI context when you put a new signature dish on the menu.
  • Voice drift. When the team finds itself saying something three times to guests, write it down once so the AI says it too.

What happens behind the scenes

When you save your venue context page (or a dish’s internal notes), the system splits the text by ## headings into short chunks, builds a small mathematical fingerprint of each chunk, and stores it. When a customer or waiter asks the AI a question, the AI finds the chunks with fingerprints closest to the question and reads them.

This means the more specific your writing, the better the answer. A vague paragraph (“we use good ingredients”) gets matched to vague questions and produces vague answers. A specific paragraph (“our 00 flour comes from Mulino Marino in Cossano Belbo, milled on the day we bake”) will be pulled up the moment any flour, bread, or pizza question arrives.

It also means you don’t have to repeat yourself. The AI reads your flour supplier directly from the inventory item, your pizza recipe directly from the recipe, your tonight’s price directly from the till. Don’t restate facts that are already structured somewhere else — write only the why, the story, and the voice.

For the full picture of which information lives where, see Where your information lives — five places, one restaurant.

Examples

  • 🍕 Pizza pairing question — A diner asks “what beer for the diavola?” The AI reads: (1) what beers you have available tonight, (2) that diavola contains spicy salami, and (3) your house entry that says “we like German hefeweizen with our spicy pizzas — the wheat softens the heat.” The diner gets a specific recommendation grounded in your kitchen, not a textbook.
  • 🍅 Allergen reassurance — A guest asks “is the tomato sauce gluten-free?” The AI quotes your allergen explanation and your sourcing entry, but never confirms safety alone — it escalates to the manager. This rule is built in.
  • ⚠️ Edge case — A waiter asks “what’s the difference between guanciale and pancetta?” If you have a voice entry that says “we say guanciale; some places use pancetta but it’s a different cut,” the AI uses your wording. If you don’t, the AI falls back to the shared knowledge file and gives a generic answer. Writing your house preference once is enough.