# AI transparency

Two features use a model: reading a recipe photo and matching ingredients. The plan and the basket are arithmetic. This is what goes in and what does not.

Last reviewed: 2026-11-07

## Where Cuadra uses AI

Cuadra uses a language model in two specific features: reading a recipe photo you upload, and matching ingredient names against the canonical catalog. This English text is a courtesy translation; the Spanish version is the operative one.

The planner that assigns dinners, calculates quantities and totals the basket is deterministic arithmetic. We do not call it AI and it does not use a model.

## Which model and who runs it

We use Google Gemini through the @google/genai SDK. The deployment can run in two ways: with a Gemini API key or with Vertex AI, which uses a Google Cloud project and the configured region (eu by default).

If the deployment has no credentials, the photo feature reports that it is unavailable; the rest of the service keeps working.

## What we send

For photo reading we send the image, its MIME type and the recipe language. For matching we send the ingredient names and the canonical catalog. We do not send your email, account, plans, basket or allergens.

A photo can contain whatever you choose to photograph, including other people’s data. The image travels in the same request and is not stored on our servers; the result is kept only if you save it as a recipe.

## Training

On the paid service, the Gemini API additional terms state that Google does not use prompts or responses to train its products. On the free tier, the terms allow their use to improve its services.

That is why the key must be confirmed as paid tier before enabling the feature in production. If it is not, the feature must not be switched on.

## Limits

The model can be wrong when reading a photo or matching an ingredient. Always review what was extracted before saving. Cuadra is not a medical or nutrition service and does not replace a professional.

Deterministic matching runs first; the model only suggests when the rule finds no candidate, and its answers are validated against the catalog. Suggestions below 0.8 confidence are discarded.

## Human oversight

You confirm: the extracted recipe is shown for review and saved only if you save it. Unmatched ingredients go to a review queue; they are never created automatically.

There are no automated decisions about your account, access or price.

## Rights and complaints

If an output looks wrong or harmful, write to contract@murabi.io with the plan or recipe and we will review it. You can also complain to the AEPD or the Estonian Data Protection Inspectorate.

The personal data processed around the model is described in the privacy policy.

[Read the privacy policy](/legal/privacidad)

## Content marking

Article 50(2) of Regulation (EU) 2024/1689 requires synthetic content to be marked. Today we do not generate synthetic text, images, audio or video: we extract text from your photo and match names. When we add content generation, we will mark it in a machine-readable way.

There is no conversational surface either. If we add a chat, we will disclose inside the interface, at the first interaction, that it is an AI system.

## Changes

Every version carries a date and a number. This page changes when the model use changes.

---
Source: https://cuadra-staging.murabi.io/en/legal/transparencia-ia