# How estimates work

> Where the number comes from — and why Kallo starts from text rather than a photo.

You do not need this page to use Kallo. But if you want to know where the number
came from — and how much to trust it — here it is.

There are two things worth knowing, and both are things you can verify on your
own screen.

## The numbers come from data, not from the model's memory

**The AI model does not invent the nutrition figures.** It does reading
comprehension: separating your sentence into dishes, then into ingredients, with
the cooking method attached. The numbers come from food composition data — the
[Vietnam National Food Composition Table 2007](https://www.fao.org/fileadmin/templates/food_composition/documents/pdf/VTN_FCT_2007.pdf), plus [FAO](https://www.fao.org/food-composition/tables-and-databases/en/) and [USDA](https://fdc.nal.usda.gov/) data.

That is why `chicken caesar salad` comes back as four components, each with its
own row of numbers, rather than one figure for the whole plate.

> **If you write in Vietnamese**
>
> Kallo does not strip accents from either the query or the reference data,
> because `bò` (beef), `bơ` (butter) and `bổ` (nutritious) would collapse into one
> if it did. Writing with diacritics gives markedly better results.

## When there isn't enough to go on, Kallo asks

The thing that decides most of the accuracy is not recognising the dish — it is
the weight. "One bowl" could be 200ml or 700ml, and that gap carries through
every number after it.

If you stated a weight, Kallo uses exactly that. If the dish is genuinely too
ambiguous to settle, Kallo **stops and asks which size you used**, with
reference photographs — rather than picking something in the middle and
presenting it as though it were certain.

> A useful estimate, not manufactured certainty.

## Why text, not a photo

Kallo starts from the sentence you write. That is a deliberate choice, not a
photo feature we haven't got to yet.

|                        | Describing it in text            | Taking a photo                    |
| ---------------------- | -------------------------------- | --------------------------------- |
| What you can see       | Every ingredient, every number   | One number for the whole plate    |
| Fixing one detail      | Correct that exact detail        | Reshoot, re-guess                 |
| What a lens can't see  | Oil, broth, sugar, what's buried | No way to know                    |
| Logging the same meal again | Same sentence, same result  | Lighting and angle change it      |
| Speed                  | You have to type                 | Faster                            |

That last row is the real cost, and we are not hiding it: typing a sentence is
slower than holding up a camera. In exchange you get what a photograph cannot
give you — you see what Kallo understood, you can correct the exact thing that
is wrong, and logging the same meal tomorrow gives the same number.

Put another way, this is **manual logging done properly**: you still decide what
you ate, you just don't have to translate it into ten separate searches.

> **The real limit of this approach**
>
> A description is only as good as what you actually know about the dish. With
> home cooking, you know how much oil went in. With restaurant food nobody does —
> and no logging method solves that. Detail in
> [Accuracy and limits](https://kallo.fit/en/docs/estimates/accuracy-and-limits).

## What is never sent

The model receives only your meal description, your countries, your cooking
habits and the output language. Your goal, targets, weight, age and sex stay on
the server. Detail in [Your data](https://kallo.fit/en/docs/account/your-data).

## Read next

[Accuracy and limits](https://kallo.fit/en/docs/estimates/accuracy-and-limits) covers where this
approach goes wrong, and by how much.

---

_Source: [https://kallo.fit/en/docs/estimates/how-it-works](https://kallo.fit/en/docs/estimates/how-it-works) · Last updated: 2026-08-03 · Kallo docs index: [https://kallo.fit/llms.txt](https://kallo.fit/llms.txt)_