A language model alone produces plausible output. iAlacarte surrounds it with deterministic layers: the menu is extracted then checked by rules, every question is first resolved through filtered search, the model writes only from selected dishes, and its answer is reviewed by guards before being delivered.
A generic chatbot answers from memory: it knows neither your menu, nor your prices, nor your opening hours. At iAlacarte, the language model receives a short list of already filtered dishes for each answer, with their data, and is not allowed to mention any others. Factual questions do not even need it.
The guest's question is understood (intent, restrictions, budget), then dishes are searched within the served menu using hard constraints: allergens, diets, excluded ingredients, calorie or price cap, current service. The language model writes only from these candidates.
Factual questions — “what is in this dish?”, “how much does it cost?”, “what time do you close?” — receive an answer written directly from the data, without a language model.
Excluded dishes are named; the precaution about traces is recalled once per conversation; confirmation is made with the team.
A register of sourced criteria (alcohol, including cooked alcohol, raw food, cured meat, raw milk, mercury…) excludes dishes; a requested wine is discouraged.
Diabetes, salt, potassium, cholesterol, histamine: indicators come from sourced criteria and estimates, never from a diagnosis.
No personalised advice; the guest is referred to their doctor or pharmacist.
From the dashboard, you can correct a name, ingredients, a price, a photo, or a section title. The correction applies to subsequent answers and is retained: it is reapplied if the menu is analysed again. The “Check the menu” screen flags what the guest would see incorrectly. Only one menu is served per location, and you decide which one.
For the legal framework, see Terms of Use & limits of liability.
Import your menu: the “Check the menu” screen shows what the guest will see and what still needs correcting.