Feature: Allow user to select model (GPT, o1, o3, etc) per chat

[orngeatom]

Not an issue but a request to implement an azure chat model selector that can be selected at start of chat

[olafwrieden]

Nice idea! There are different capabilities in these models so for example some might do audio and image whereas others are purely text-based - some redesign / disabling features based on selected model may need to happen. You could put a model selector on the chat text box which is probably the best place as it's consistent between personas/extensions/chat pages and would require minimal changes to UI.

A challenge I see is deploying a multi-model approach with infra as code. This was the case with DallE which wasn't available in all Azure regions, meaning perhaps we take a GPT4o model as baseline for the IaC deployment, then allow deployments from AI Foundry to be configured via an array similar to how we handle configuration for Identity Providers. If a "custom" model deployment has been provided, the model selector becomes visible allowing the user to select the out of the box 4o or configured custom deployment.

Now we would just have to handle changes to the request body and function calling based on model schema and capabilities - maybe something Semantic Kernel could help with.

Example of Model Selector

[fsch-ppi]

I have a very simple implementation with a model slider that you can change when on a chat. I have extended the ChatThreadModel with the used model, the api call is then done to a respective endpoint depending on this attribute.

[fsch-ppi]

Short explanation for the dalle3 slider: Since the chatbot will quite often generate images even when not asked about it.
I have changed the code in a way, so that the ability to create images is only explicitly given to the chatbot, when the dalle3 slider is selected (gpt4o is the underlying model in our case).

As soon as MS unlocks us for the o3 model for an EU deployment we will also integrate that into slider and chat.

[orngeatom]

@fsch-ppi this is along the lines of the goal.... if i can get more information on thi. based on this i would say have additional variables connected to Azure Open AI models and the card lets you essentially trigger which model to use. All other functions remain the same , very great! working to implement this... @olafwrieden this should be part of this repo if it wants to live in the waters of all other repos . we can similar to the azure open ai model set though. happy to share what i can find here..

buttons can be added to the chat-header.tsx, a new script can be created to design the buttons the same way the others are... the problem im having is that the client gets updated with the button but i can't pass the button value to trigger another config. any ideas without changing architecture on how to trigger an function based on the client pass to server get ?

[fsch-ppi]

For each ChatThread I added a "usedModel" attribute as a string, the default one is o4-mini for our company. This will however require a database update for each change. However they are rare, as the user changes model once per interaction with a Chat Thread I would say.

You could also store it in the browsers local store and have a mapping from chatThread to the usedModel, then make a call based on the variables in the users local store. This will grow per user usage and we deemed it had no advantages for us.