Add a new text command for a more ergonomic way to attach text to the chat interface #1476
Conversation
|
@joadoumie Thank you for contributing this! I just got back from some software conferences, so I appreciate your patience until I'm able to review this. BTW, we have recently added a feature where users can drag & drop selections into the chat input. That might solve the original issue for this PR, but I'm also open to having these 2 features live together in parallel. Have to test this locally to arrive at a recommendation. |
|
I think drag and drop is a great feature, and probably solves for a different user flow (folks who prefer interactivity with the keyboard and keyboard shortcuts). I'd probably use both, but it'd be interesting to try them out together. |
|
I think it would be helpful to be able to attach using keyboard shortcuts, but we should use the new attachment system for the content. This is a more flexible and powerful way of attaching context from different documents to the chat that gives agents/personas for control over what appears in the context/chat history and has a more streamlined UX. |
There was a problem hiding this comment.
Thank you for contributing this! This is a great addition to Jupyter AI & provides a keyboard-only option for adding a selection as context. There are 2 potential improvements that I'd like you to explore before merging:
- The added context should be attached as an attachment instead of being injected directly into the prompt. This can be done via
inputModel.addAttachment(). A reference for this can be found inside of the@filecommand provider. The attachments API should now support selection ranges.
- Thanks to your previous contribution in #1419, this attachment will be read by Jupyternaut. Right now I believe it will read the whole file. There's no need to implement reading only selection ranges in this PR since we will add an agent tool for this (read further).
- A user generally only has one selection at any given time, so I recommend just providing a single
@selectioncommand instead of two@textcommands. If a text selection is available, use that, otherwise use the cell selection.
I totally understand any reservation about using the attachments API. It is more complex and the UX can be improved, but there is a good motivation for this. By v3.0.0 official, Jupyternaut will be made agentic & be provided tools to read files on their own. I have a local branch of this working and will open a PR for this next week.
By having the AI agent load the content itself, we can reduce the time-to-first-response by having a smaller initial prompt, and most importantly, make AI reads/writes to the filesystem traceable for server operators & admins.
(congrats on getting feedback from the co-founder of Project Jupyter BTW! that's Brian.)
|
@dlqqq -- I made some changes given your feedback. I don't think the attachment API quite fits the desired UX just yet, but I think this is directionally correct. I also took a look at #1479 to try to better understand how this flow may fit in with a world of agentic tool calling (which I'm really excited about). I am going to temporarily move this PR into draft status so I can revisit it after better understanding some of the possibilities/limitations with the tools based approach. |
Description
This PR introduces a new TextCommandProvider that allows users to include text content directly into chat messages via
@text:commands. This allows users to quickly include active cell content or selected text in their chat messages, providing an alternative to the dropdown menu interface. This is a more ergonomic method of attaching commands for folks that use keyboard over mouse.Code changes
New File
packages/jupyter-ai/src/chat-command-plugins/text-command.tsxTextCommandProviderclass, following the same pattern asFileCommandProvider.@text:active_cell– inserts the content of the currently active notebook cell.@text:selected_text– inserts any text currently selected in the editor.actonly shows theactive_celloption).Modified
packages/jupyter-ai/src/chat-command-plugins/index.tstextCommandPluginto the exported plugins array.User-facing changes
Users can now type
@text:in the chat input to trigger autocomplete suggestions:@text:active_cell→ Includes the content of the currently active notebook cell.@text:selected_text→ Includes any text currently selected in the editor.Backwards-incompatible changes
None that I know of.