Increase AI text output field in larger increments instead of line by line

mhersh

When using Quick Answer or the assistant, Kagi expands the LLM output line by line. This makes it difficult to use the page while the LLM is processing.

In search results using Quick Answer, it continuously pushes the search results down as the quick answer area expands, so I cannot read and use the static search results below while I wait for the LLM. This is especially disruptive when using the !quick bang (or ending my query with a question mark), because it effectively changes the results loading time from a few milliseconds to a few seconds or more.

In the assistant, by default it automatically scrolls down line by line, which makes it very difficult to read the response until it finishes. You can work around this by scrolling up slightly, but as soon as you scroll back to the bottom it becomes jumpy again.

Ideally, pages should be easy to read and interact with while the LLM is processing by default.

I have a few suggestions for how this might be improved:

Instead of expanding the results line-by-line, expand it in 5-line or 10-line chunks, leaving enough empty space that it will not need to expand again so rapidly. This will still be a little jumpy, but it would improve readability without major changes in design.
Move the Quick Answer field into the sidebar when available, so it will not interfere with the placement of search results. (Would only help on desktop.)
Disable auto-scroll in the assistant, at least as an option. I am perfectly happy to do my own scrolling, and I don't find auto-scrolling helpful.

Vlad

mhersh You can already select "Show less" option and Quick Answer output will be constrained and fixed in height until you expand it.

mhersh

That's true, but there are a few problems with this:

It shows only an extremely limited amount of the response, to the point that it is not generally useful until expanded.
The "Show Less" button is itself a moving target, just like the search results, so it's very difficult to use until generation is complete. This should probably be moved to a stable position at the top or side of the response text.
I don't think there's a way to have it always open in collapsed view. It seems to remember the last state, which will likely be expanded (since of course I want to read the result, so I will expand it eventually).
When collapsed, I don't see any indication of when the reply is completed.

I'd rather not choose between usable search results and a usable quick summary. Ideally, both would be readable and easily targetable from start to finish.

Vlad

mhersh The idea is that you would expand it and collapse it when done. The button is not a moving target in collapsed move, it stays in fixed position. We introduced it as a result of feedback like yours and people keep it in collapsed mode by default, expanding when needed. How would you change this?

mhersh

As a low-effort, low-impact change, I would add a setting with a choice between three options: 1) Always start collapsed, 2) always start expanded, and 3) remember last state.

For me personally, the last state is simply not a valid indication of how I want future searches to behave. To me, remembering this only introduces inconsistency and unpredictability. Manually re-collapsing it before closing the search results page is an extra step that could be avoided. My workflow in search results pages is usually to open links from them and move on. If I'm opening links in the same tab, I will have no opportunity to manually re-collapse the view. If I am opening links in new tabs (my default), I don't want to have to remember to go back to the results tab and collapse it when I'm done. I am not realistically going to put my actual tasks on hold to spend time micromanaging my settings with every search.

That said, I wouldn't consider a new setting a full solution, because it does not address the usability problem when it's expanded, or in the Assistant. The problem there is simply stated: rapidly moving text and links are unusable. Why should the majority of the loaded content on the page be unusable while the LLM is still working?

My previous suggestion of making the text box expand in larger chunks instead of line-by-line would help mitigate this problem (though not entirely solve it). Alternatively, perhaps there could be a middle ground between "expanded" and "collapsed", where the box is statically-sized but large enough to use, with its own scroll bar. Maybe 10 lines tall or something like that. Then you could eliminate the jumpiness altogether, and both the quick answer AND the search results would be usable at the same time immediately upon loading.

As for the assistant, auto-scrolling line by line adds a little extra friction every time I submit a query. I can scroll up and stabilize it, but I feel like I need to fight against the UI every time. Again, the fundamental problem is that jumpy text is not useful. Jumpiness should be minimized if it can't be eliminated.

Thanks for your attention. I'd be happy to discuss this further, and I hope you and your team can come up with better solutions to the fundamental problem than I can off the top of my head.