Sovereignty is described by levels

“Sovereign” has become a common word. It is often used as a slogan. We want to use it as a description.

The problem with “fully sovereign”

A data and AI solution is not a single block. It stacks several layers. Each layer can be local, or not.

Hardware is rarely made in Morocco. Much open source software is built by international communities. The most capable language models are often offered by foreign providers.

Saying “fully sovereign” hides all of this. It is not honest, and it does not help to decide.

Six levels

We describe a solution in six levels:

1. Storage. Where is the data?
2. Query engine. Where are the computations done?
3. Catalog and permissions. Who describes the data, and who decides on access?
4. Embeddings. Which models turn text into vectors for search?
5. Language model. Which model reads and answers, and where does it run?
6. Operations. Who monitors, updates and secures day to day?

For each level, we answer a single question: is it local, yes or no, and why?

An example

An organization can keep its data in Morocco, with an open engine and a locally hosted catalog. The first three levels are then local.

It can also choose an external language model, through an API, because the expected quality requires it. That level is not local. We have to say so, and control what is sent to it.

This solution is not “fully sovereign”. It is described precisely, and everyone can judge.

What we will test

Two levels interest us in particular:

• the quality of open embedding models in Darija and Amazigh,
• the real cost of an open language model hosted locally.

We will publish our measurements here, with the method.

Our detailed position is on the Sovereignty page.