This is a proof of concept testing Transformer.js with Supabase Edge Functions. Executes the edge-runtime directly via Docker.

npm run start

This runs a docker compose command under the hood, listens on port 9000.

We have a function named generate-embeddings that lives under ./supabase/functions/generate-embeddings. It loads the all-MiniLM-L6-v2 model from Hugging Face and generates a 384 dimension embedding vector based on the a JSON input string passed to the function (via POST request):

{
  "input": "The cat chases the mouse"
}

You can test the function using cURL:

curl -i --location --request POST 'http://localhost:9000/generate-embeddings' \
  --header 'Content-Type: application/json' \
  --data '{"input":"The cat chases the mouse"}'

Transcribe audio from a WAV file using Whisper.

There is a function named transcribe that lives under ./supabase/functions/transcribe. It loads the whisper-tiny.en model from Hugging Face and transcribes a WAV file passed to the function via multi-part form data in a POST request:

curl -i --location --request POST 'http://localhost:9000/transcribe' \
  -F audio=@"my-file.wav"

You can monitor the container's resource usage statistics in realtime using docker stats:

docker stats edge-runtime