Wrong AI answer to : "what is your AI and LLM ? / reply **GPT‑4** d’OpenAI.

[Vincentparisgithub]

Bug Report

Before you file your issue

• Check the other issues before filing your own
• If your report is related to the (BlockNote) text editor, file it on their repo. If you're not sure whether your issue is with BlockNote or Docs, file it on our repo: if we support it, we'll backport it upstream ourselves 😊, otherwise we'll ask you to do so.

Problematic behavior
When we ask to the AI "what is yourAI and LLM"

the AI answers : GPT‑4 d’OpenAI.

Expected behavior/code
the AI should answer the right answer :
https://docs.numerique.gouv.fr/docs/ae3d212d-1bf6-4d67-a524-b34fba59ca51/

Quelle IA est utilisée ?

L'IA intégrée à Docs est mise à disposition par [Albert API](https://alliance.numerique.gouv.fr/albert/), un service développé par la [DINUM](https://alliance.numerique.gouv.fr/albert/). Actuellement le modèle utilisé est [GPT-OSS-120B](https://comparia.beta.gouv.fr/modeles#gpt-oss-120b).

Albert API est opéré par la DINUM et hébergé en SecNumCloud. Le contenu des requêtes ne sont pas stockés par Albert API pour éviter d'exposer des données sensibles.

Vos requêtes ne servent à pas à ré-entrainer des modèles.

Côté Docs nous utilisons [Langfuse](https://github.com/langfuse/langfuse), un outil open source que nous auto-herbergeons, qui nous permet de d'analyser la qualité des réponses du LLM.

Steps to Reproduce

1. Do this...
2. Then this...
3. And then the bug happens!

Environment

• Docs version:
• Instance url:

Possible Solution

Additional context/Screenshots
Add any other context about the problem here. If applicable, add screenshots to help explain.

[AntoLC]

The root problem is not on Docs side, Docs is "LLM agnostic", any llm can be plugged.
The issue is on Albert side, the model should probably have more training on this topic.
If you do the same request on the playground, you will probably get the same answer.
https://albert.playground.etalab.gouv.fr/

That said, we can maybe find a solution on Docs side.

[virgile-dev]

@AntoLC like Eliott said we should probably have a system prompt so that the answer is correct.
Most self hosted LLM are not re-trained so Albert API won't solve it. But the OpenAI team training GPT-OSS could do smthg. I happen to know the PM. I'll send him the issue.