Option for auto-learning project root

[dlqqq]

Problem

Some users would like to automatically invoke /learn on their entire project root at server start to make local RAG easier.

Proposed Solution

Right now, I think whether this should be implemented is still an open discussion. I personally am against this, as it violates our principle of never implicitly invoking upstream models without an explicit user action.

[liyongsea]

Hi, our team has been working on the jupytercoder as a chrome extension. It would be amazing if we could integrate this natively into the jupyter notebook. Starcoder has been trained on purpose on some notebook data to incoporate the output of the cell.
Do you think this will fit into the principle that dlqqq mention above ? @JasonWeill
Otherwise we are very happy to start trying to implement this

[dlqqq]

@liyongsea Could you offer some more details on what you want your PR to implement? We can provide helpful guidance to help navigate your team through our code.

[liyongsea]

It is basicly a codex link plugin for jupyter notebook. Currently it is implemented as a chrome extension. We use a trick to simulate the input of a user by modifying the DOM
Here is a demo of the expected behavior https://twitter.com/BigCodeProject/status/1654174953565044736
Here is what we want to do:

• read the content of the cells, including the output
• send to starcoder, which has been trained specifically on notebook
• suggest to result as grey text
  I would like to first validate the interest of such a plugin

[dlqqq]

@liyongsea We are most certainly open to the idea, and would welcome a PR. But this feature would have to live inside the current labextension code and not as a separate Chrome extension to be used with Jupyter AI. I expect this to be a fairly difficult task, as you would have to use some JupyterLab autocomplete API in addition to streaming LLM responses from the Jupyter AI server extension.

I also wanted to emphasize that this is not something we wish to offer by default. We would never want to automatically send the contents of a user's notebook to a remote LLM without their explicit consent.

[liyongsea]

Hi @dlqqq thank you for your reply !

But this feature would have to live inside the current labextension code and not as a separate Chrome extension to be used with Jupyter AI

This is also our will, because with the chrome extension, we modify directly the DOM, which is very difficult to maintain

I expect this to be a fairly difficult task, as you would have to use some JupyterLab autocomplete API in addition to streaming LLM responses from the Jupyter AI server extension.

It would be nice if you could point us to a starting point. The difficult part in our previous experience, is to find an appropriate way to modify the input cell in the notebook

I also wanted to emphasize that this is not something we wish to offer by default

Totally agree.

[dlqqq]

@liyongsea I would highly recommend two resources:

1. The JupyterLab extension development guide

2. The JupyterLab JavaScript API documentation

The way to do pretty much anything client-side is via the JupyterLab JavaScript API, which unfortunately isn't that well-documented. I think the best way to learn extension development is to look at our source code. Keep in mind the JavaScript entry point is almost always under src/index.tsx for any labextension.

[liyongsea]

Hi @dlqqq ,
We have made some progress. This is a functionning prototype.
Here is the main idea:

• add a new side bar in https://github.com/jupyterlab/jupyter-ai/blob/main/packages/jupyter-ai/src/index.ts to configurate the autocomplete fonctionality
• add keydown event listen (in a frist version, user has to type something to trigger the autocomplete. Will improve it very soon)
  It would be part of the current jupyter ai. User will have to activate the plugin. In a first version, user will use a key to trigger the auto complete.
  If you think this make sense, we will create a PR. Is there any coding style/testing best practice for us to respect ? Who do you think is the bes person to review the PR ?

[liyongsea]

Hi @dlqqq
We have built a functionning prototype, shall I create a PR for you to review ?

[dlqqq]

@liyongsea Please feel free to submit a PR! We welcome all contributions. I can help with the review.

[liyongsea]

Hi @dlqqq #378 here you go :)