Photo, voice, label, recipe: where your meal values come from

The principle

A food diary is only useful if its figures are reliable. Azetta therefore clearly distinguishes three sources: values measured by laboratories and published in official tables, values read from packaging, and values estimated by artificial intelligence. Each way of adding a meal relies on one of them, and the app tells you which.

The Azetta method

The food database: official tables

The built-in database contains 1692 foods: 942 from the American USDA SR Legacy table and 750 from the French CIQUAL 2020 table (ANSES). For each one, the values per 100 g — calories, protein, carbohydrates, fat, fibre, sugars, saturated fat, sodium — are copied from the tables, never written by hand. A serving is calculated simply:

For liquids, 1 ml counts as 1 g. The salt shown comes from sodium: salt = sodium × 2.5. A scanned barcode queries Open Food Facts; if the product only gives kilojoules, calories equal kJ ÷ 4.184.

A meal spoken or typed: the database first

When you describe your meal ("two eggs and a slice of toast"), the AI splits the sentence into foods and quantities. Then the app looks up each food in the database: if it is there, the official values are used. Only foods that can't be found keep the AI's estimate, marked "Estimate". You can also choose the coach's estimate for a line. The server brings impossible answers back within bounds: at most 9 kcal per gram, and no macro heavier than the serving itself.

A label: read, never invented

For a nutrition label, the AI reads the printed values; an unreadable value stays empty. The server then converts: kilojoules to calories (÷ 4.184), sodium to salt (× 2.5), values per serving to values per 100 g. It refuses impossible values (more than 900 kcal or more than 100 g of a nutrient per 100 g) and flags an inconsistent reading, for example if sugars exceed carbohydrates, or if calories differ by more than 15% (and by at least 20 kcal) from 4 × protein + 4 × carbohydrates + 9 × fat + 2 × fibre.

A photo: an estimate

From a photo of your plate, the AI estimates everything: the foods, the grams, the calories and the macros. It is the fastest and least accurate method, and the app says so: "Estimates from the photo: adjust the grams if needed." If you correct the grams, all the values on the line follow proportionally.

A recipe: the sum of the ingredients

The weight of the dish is the cooked weight if you give it, otherwise the sum of the raw ingredients. A detail (sugars, fibre…) is only shown if all the ingredients provide it.

Why this choice

• The USDA and CIQUAL tables are public, documented and based on analyses: they are the safest possible basis.
• An AI is good at understanding a sentence or reading packaging, much less at guessing a weight from a photo. Hence the rule: it understands, the database supplies the numbers; it never fills in a label.
• The conversion salt = sodium × 2.5 and the energy factors are those of the European labelling regulation.

Limits

• A photo sees neither the cooking oil nor what is hidden under the sauce: treat the result as an order of magnitude and correct the grams.
• Matching with the database is done by name: "bread" will find a bread, not necessarily yours. Check the suggested line.
• A recipe's values assume you weighed the ingredients.
• The AI features have a daily usage limit.

References

1. U.S. Department of Agriculture, Agricultural Research Service. USDA National Nutrient Database for Standard Reference, Legacy Release (SR Legacy), 2018.
2. ANSES. Ciqual 2020 food composition table.
3. Regulation (EU) No 1169/2011 of the European Parliament and of the Council of 25 October 2011 on the provision of food information to consumers (Annex XIV: conversion factors).