For long AI replies the current UI is not fit

[virgile-dev]

What’s broken?

One of our power users gave a legit feedback:

"Avoid simulating streaming; it is too time-consuming for large texts."

When the AI reply is very long (generation or select) the current UI where you can see the AI cursor typing text in the editor is not fit as it takes too long to get the result.
Sure you can adapt a the delay but maybe it won't be confortable for the users.

What did you expect to happen?

When the AI has to update / generate large chunk of text it could be nice to output it all at once or by large chunks.

Steps to reproduce

I've done a bit of benchmarking betwen Notion and BlockNote AI to show the difference between the approches on several use case :

• translate a large text in english into french
• make a large text shorter
• continue writing

Always using the same text sample
https://fichiers.numerique.gouv.fr/explorer/items/2dadfbd6-170b-45da-ab42-d15832e9f978

BlockNote version

No response

Environment

No response

Additional context

No response

Contribution

• I'd be interested in contributing a fix for this issue

Sponsor

• I'm a sponsor and would appreciate if you could look into this sooner than later 💖

[virgile-dev]

Also related to large text selection, you might lose sight of where the AI tool bar is.
Notion has a cool fix for this they add a icon at the bottom or the top that you can click to scroll to where the AI tool actually is.

[virgile-dev]

Also an interesting feedback Antho gave me is that the current UI is not adapted for some task.
For example in the context of a translation it doesn't make sense to display suggestions.
Here is how Notion deals with it, by making a proposal that the user can insert below.
https://fichiers.numerique.gouv.fr/explorer/items/files/302a13be-8e49-42c6-b286-7f6a97705651

[canal-side-coder]

I too find the simulated streaming is impacting user experience.

I am currently evalutating blocknote AI as a solution for a commercial project. This issue has quite a high impact straight away. Especially in scenarios where extended thinking or long running tasks may have already taken some time before the first token is outputted then adding additional delays becomes noticeable.

I have started experimenting with Remend in order to stream markdown directly as the response output, similar to how Streamdown behaves. It would be great to hear your thoughts on this approach.

[YousefED]

We'd love to explore different UX options (both smarter animation timings and different interfaces such as side-by-side compare), hopefully later this year (or earlier if there's a sponsor for this).

Note that for now, it's possible to disable the delays completely (see [withDelays](https://www.blocknotejs.org/docs/features/ai/reference#getstreamtoolsprovider))

@canal-side-coder if you have other (commercial) requirements, always happy to collaborate - reach out to team@blocknotejs.org