Background

There was a fantastic idea from a Redditor about making EnergeticAI run well in Cloudflare Workers, so you can have super fast inference at the edge, without the need to distribute your model weights:

https://www.reddit.com/r/tensorflow/comments/1493uoq/comment/jo6axc9/?utm_source=reddit&utm_medium=web2x&context=3

Goal

This task is to add support for Cloudflare Workers to EnergeticAI.

Approach

Given that Cloudflare Workers have even more restrictive bundle limits than AWS Lambda, I suspect the way to do this would be to distribute sharded model weights in Cloudflare KV, and then fetch from that in parallel on function invocation. On paper at least KV values should be colocated with the functions enough that this should be fast.

EnergeticAI currently uses the lite Universal Sentence Encoder model, which is English only:
https://tfhub.dev/google/universal-sentence-encoder-lite/2

It's possible we could convert one of the smaller multilingual models on TFHub and explore if there's a way to get it to be usable:
https://tfhub.dev/google/universal-sentence-encoder-multilingual/3

Background

There was a fantastic idea from a Redditor about making EnergeticAI run well in Cloudflare Workers, so you can have super fast inference at the edge, without the need to distribute your model weights:

https://www.reddit.com/r/tensorflow/comments/1493uoq/comment/jo6axc9/?utm_source=reddit&utm_medium=web2x&context=3

Goal

This task is to add support for Deno Deploy to EnergeticAI.

Approach

Given that Deno Deploy has even more restrictive bundle limits than AWS Lambda, I suspect the way to do this would be to distribute sharded model weights in Deno KV, and then fetch from that in parallel on function invocation. On paper at least KV values should be colocated with the functions enough that this should be fast.