What was the most surprising thing you saw or learned in this rewrite process?
The move from Electron helped us remove "native libraries" in Node. Previously we were using Xenova transformers to instantiate the models. This forced us to use multiple package.jsons and added a lot of cognitive overhead to our development. The move to rust freed us up to focus more on feature development than configs and setup. It was surprising because I thought learning rust would set us back much longer, but th…
Would have loved to use sqlite if I could get good results. Maybe I botched the implementation.
Creator of sqlite-vec here, happy to help debug anything if you do attempt to try out SQLite vector search again. You mentioned "VSS" in another comment, which if that was my sqlite-vss extension, I more-or-less abandoned that extension for sqlite-vec last year. The VSS extension was more unstable and bloated (based on Faiss), but the new VEC extension is much more lightweight and easy to use (custom C vector functio…
Ended up swapping out redis for sqlite-vec. Works great. Thanks
I'm curious how you're implementing the image similarity matching. I recently reverse engineered the Apple Neural Hash model and wrote an API to use it in my app for doing image similarity calculations. I found it to be extremely quick compared to some of the other more computationally intensive methods that I was trying to use before.
We're using redis vector modules for cosine similarity. I'm sure there's more to optimize there. Your project sounds cool. How'd you reverse engineer apple's model?
I extracted the espresso model files from a recent iOS build and then wrote a python script to unpack the weights and save them in an open source format that is easier to work with, in my case it was onnx. Through many rounds of iteration, I converted each neural network layer from espresso to its onnx equivalent and eventually got something that works, then turned it into an API. If you want me to send it to you, let me know! My GitHub username is the same username I have on here, and on my GitHub you’ll find my email.
Please note that Redis 8 supports vector sets as a native type without building any module. And the API is very simple to use. This may simplify the building / bundling process perhaps.
> Please just let us bundle a modern chrome with our apps. please, no. I wish software companies had to pay the hardware they require for their users, then we would have devs using Rust instead of JS and optimizing using ASM just to save parts of cents per instance. And we wouldn't see companies like MS kill well designed and performed native apps for a electron app
Why are you against the idea of an optional feature that lets them keep using Tauri while maintaining browser parity? You don't have to use it.
Besides it's a crime against the environment: The user doesn't have a choice