Added model selector, progress bars, and a worker.

[varunneal]

As discussed, we should have an option to load any model. This feature is now in advanced settings. Additionally, drawing from Xenova's transformers.js demo, we have a loading bar corresponding to downloading the models. This requires using workers, which involved some code refactoring.

As always, the demo for this PR is live at https://varunnsrivastava.github.io/SemanticFinder/.

A few key points:

• It is critical we replace a similarity threshold with a top-k results threshold instead. Testing the other models makes this clear, since the spatial geometry of embedding vectors (& thus the cosine similarity) varies greatly between model. I haven't made this change live yet, but I can after some discussion.
• I've replaced the previous progress bar with a Bootstrap Progress Bar
• In advanced settings we can add a toggle to enable/disable browser caching (it's on by default), since some of the larger models can take up a lot of space.
• I've begun some work benchmarking WebML embedding models. Here's a spreadsheet I hope to update frequently. Let me know if you'd like to help out, have any suggestions, be added as a collaborator, etc. I can also provide some boilerplate code I'm using for benchmarking.

[varunneal]

According to MTEB, I'll probably go ahead and convert instructor-large and probably e5-small-v2 to ONNX weights and host them on HF sometime soon.

[do-me]

Thanks a lot for this awesome feature, that's some great work! I'll have a look at your benchmarks later, I'm very curious about the results.

Before merging, I think I found two little things we should correct first:

1. Especially with other models, white spaces are "polluting" the results. In this example (with the settings displayed), all top results are white spaces strangely.

2. I noted, that if you run the same query two times (in this example clicking Submit a second time after the first run) it will re-append the results, leading in this example to even more white space entries on the top. Can you reproduce this behavior?

I just tested e.g. the multilingual model with French, Italian and German and indeed mostly the results are quite good (for such a small model of only 135Mb!).

Yet, e.g. in this example with Italian, white spaces for some reason generally have a high similarity score and mess with the results:

3. [not crucial] By the way, I like the appearance of the progress bars but somehow they bounce around a lot :D Like when e.g. you rerun something, they take a little too much time to go back to the start so it feels a bit counterintuitive. Maybe there's some setting tweaking the responsiveness? If it was a little snappier it would already mitigate this effect I guess.

[varunneal]

Thanks for catching these bugs! Will work on them. I'll get back to you when I

• add a filter to ignore the empty string ""
• implement top-k as the similarity threshold metric (let me know if you're against this)
• fix progress bars
• fix weird re-append errors, I think mostly associated with the "split by # words" feature

[do-me]

Top-k is a great idea for replacing the default of 0.3!

By the way, would you mind adding public viewing permissions for your spreadsheet?

[varunneal]

Yes! I haven't populated it much yet. To measure chrome memory I'm using Task Manager as an approximation. Here's a JS function to measure server-side memory usage:

function getTotalMemoryUsage() {
    const memoryUsage = process.memoryUsage();
    return memoryUsage.heapUsed + memoryUsage.external + memoryUsage.arrayBuffers;
}

I'm using process.cpuUsage(); to measure inference time.

[lizozom]

@VarunNSrivastava maybe you could split this PR into two? Doing the UI changes and the embedding changes separately?

Because splitting the package out could be a useful approach for this, but we're going to get a ton of conflicts at the moment.

[lizozom]

I thought about it some more, feel free to go ahead and merge the PR! Reintroducing the package is pretty easy!

Please share your thoughts on https://github.com/do-me/SemanticFinder/issues/24 though, as it would change my implementation!

[varunneal]

Unfortunately, the UI addition of the progress bar necessitated that I use WebWorkers for model loading and embedding, so it won't be very easy to separate the two.

@do-me The latest couple commits should fix a lot of the problems.

@lizozom It should be easy to load the embeddings right at the start, though as mentioned, it could be a very slow operation at the start if the text is large.

[do-me]

Great work @VarunNSrivastava - I'm particularly excited about testing more languages and see how well they perform!

[lizozom]

Thank you for the amazing work @VarunNSrivastava ! I'll work on revising my PR.