Display specific error when models are down

[eliott07]

Bug Report

Problematic behavior
When models don't respond because LLM provider is down, the assistant loads forever and then prompts the user to start a new conversation

Expected behavior/code
When LLM provider is down, display clear message stating that the assistant is currently not working and to come back later, maybe point to a status page either of the ASsistant or of its LLM providers (like albert.api.etalab.gouv.fr/status

[robin-lecomte]

Here is a proposal.
If the assistant goes down while a user is making a request : Proposal here.
I suggest greying out the search bar and displaying a message in the center of a new conversation, including a link to the status page (platform or a dedicated page?).

[eliott07]

For the moment maybe we can link it to a page within the documentation explaining we are still in beta and that periods of unavailability are normal

We do have to think about a status page listing status of different services (LLM inference, Web search, Collections, and later connections to docs, folders...)

[eliott07]

The initial issue was for LLMs being in error but we also have to include what happens when the Mistral OCR is down and hence we can't process PDFs for the moment.

In that situation it would mean blocking only the pdf uploads we can probably use the mechanics when we blocked all document types ? Or would that mean there has to be a feature flag dynamically activated when OCR model is down ?