Bug: Model field inputs are coupled in Settings UI

[VicaYang]

Description

I want to use different models, but seems that some of them are coupled.

Reproduce

I think at the beginning there is no API_URL for inline completition models. However, after I changing the models several times, it appers, but it is coupled with the API_URl for language model.

I also think that the inline completion model does not correctly takes the values from current UI, because it is probably a bug.

Context

I would suggest to separate the model setting for each part. For the model that support OpenAI format, we can just allow users customize MODEL_NAME, API_KEY and API_BASE_URL, just like what cursor does. Of course, we can provide a lot of default configurations, such as openai/gpt-4o-mini.

[VicaYang]

BTW, I dont think OpenRoute::* is the good choice, I would prefer OpenAI::* or simply *::*, however they are not provided.

[dlqqq]

@VicaYang Thank you for reporting this! It looks like we are not managing the frontend React state correctly in the settings UI under certain usage cases like yours.

We're currently building Jupyter AI v3, and are taking your suggestions for the chat settings into account. Yes, we definitely do want to allow custom model IDs for every provider in the future; using OpenRouter :: * is confusing for users.

[johnnyxwan]

I think this issue might have direct impact with changing base API url of the completion model as well. Please check with #1238

[VicaYang]

Yes, I think the settings of these models are highly coupled and probably cause some bugs. The configuration of API_Keys is also confusing. A code refactoring about these configurations will help a lot.

It is very unclear which URL, model, and keys are used for each component, and it seems that we cannot configure the inline completion model easily (maybe I failed to find the correct way). Making these settings transparent and configurable (but with some default values) will be a good choice, in my opinion.

[krassowski]

So I think there is a couple of issues highlighting that we need to rework the inline completions configuration and that it needs to be separate from the chat model configuration:

• Inline completion model provider should be traitlet-configurable #781 (third most up-voted issue)
• Bug: Deepseek Coder models via OpenRouter not working for code completions #1238

[johnnyxwan]

Interestingly, the backend python code does handle the 2 configurations (at least for base url) seperately. This should be a frontend issue. This mis-match is exactly the cause of #1238

Didn't look at configuration of API key.

[krassowski]

Also, there were relatively recent PRs #1125 and #1137 touching on relevant code - can you all confirm you are running latest version?

[VicaYang]

Below are the versions of all jupyter-related packages
jupyter 1.1.1
jupyter_ai 2.29.1
jupyter_ai_magics 2.29.1
jupyter_client 8.6.2
jupyter-console 6.6.3
jupyter_core 5.7.2
jupyter-events 0.10.0
jupyter-lsp 2.2.5
jupyter_server 2.14.2
jupyter_server_terminals 0.5.3
jupyterlab 4.3.5
jupyterlab_pygments 0.3.0
jupyterlab_server 2.27.3
jupyterlab_widgets 3.0.13

[dlqqq]

Thank you everybody for contributing to this issue and doing research.

The issue seems to be caused by this expression in chat-settings.tsx around line 495:

          {clmGlobalId && (
            <ModelFields
              fields={clmProvider?.fields}
              values={fields}
              onChange={setFields}
            />
          )}

The completion model fields are being stored in the same dictionary as chat model fields (namely fields). It seems like the solution would be to:

1. Rename fields and setFields to chatModelFields and setModelFields for clarity.
2. Create a new state variable to store completion model fields, e.g. const [completionModelFields, setCompletionModelFields] = useState....
3. Use completionModelFields instead of chatModelFields in the excerpt above.

Can somebody volunteer to help with this? Would love to get help from other contributors because my time is super limited right now.