Add labels for higher zoom levels

[do-me]

I recently created a large visualization with 3M points. You probably invested a lot of work in the defaults for all viz parameters as generally they produce great results.

The only thing I'd like to suggest is to add even more labels to clusters at higher zoom levels. At the moment from what I can tell you have two "levels" (5 and 10) for which you calculate labels depending on the zoom level. Here is an example where I have some very dense areas/clusters I'd like to label too.

These labels could maybe also be generated on the fly when a user zooms there.

I discovered a few hard-coded values in a svelte component, e.g. this filter:

embedding-atlas/packages/component/src/lib/embedding_view/EmbeddingViewImpl.svelte

Line 575 in 0630e2f

let threshold = maxDensity * 0.005;

Would setting this value lower to like 0.001 potentially fix my issue or should I look at other parameters?

Also, could you add Dev instructions, how to compile all the code and run it with the modifications for testing?

P.S. Could you open GitHub discussions for this repo?

[fredhohman]

Hi @do-me! We can think about whether it's worth exposing that parameter to the user @donghaoren @domoritz.

As for dev instructions, we have some here: https://apple.github.io/embedding-atlas/develop.html

Try that and see if you can make your modification work.

[do-me]

Thanks for the pointer. I installed everything with npm and built the dist but then got:

(base) ➜  backend git:(main) ✗ ./start.sh
Using CPython 3.13.5 interpreter at: /opt/homebrew/opt/python@3.13/bin/python3.13
Creating virtual environment at: .venv
error: Distribution `torch==2.5.1 @ registry+https://pypi.org/simple` can't be installed because it doesn't have a source distribution or wheel for the current platform

afaik pytorch doesn't run with Py 3.13 yet on MacOS. Not sure why this version is chosen.

Update: solved my modifying start.sh:

uv run --python 3.12 embedding-atlas spawn99/wine-reviews --text description --split train --static ../viewer/dist

But note that then the datasets package is missing from the uv file:

⚠️ Loading Hugging Face datasets requires the datasetspackage to be installed. Please runpip install datasets, then try again.

Result

let threshold = maxDensity * 0.001;

Gave me around 229 labels instead of 113 which is great! Now it's almost perfect for me. I think I'd still prefer more labels but this is a great start - would appreciate it if you could expose this parameter to the user via the CLI :)

[dataelvisliang]

Would highly appreciate if we can define the level of cluster labels, thanks a lot!!!