This is a small project intended to explore how far the in-browser LLM experience has come over the past couple of years. It uses transformers.js to download a model and run inference entirely within the browser (leveraging WebAssembly). The model used is distilgpt2 to keep bandwidth small and responses relatively fast. I might add the ability to specify the model through the UI at some point in the future - but for now, being able to change the model in a single line of code is enough for testing purposes.
I've also been enjoying Clojure recently, so I used this project as an opportunity to explore the ClojureScript ecosystem with the help of shadow-cljs. JavaScript interop was trickier than I expected - especially when importing NPM packages - but playing around with familiar React concepts in a more functional style through Reagent was a fun experience.