Running RAG with a (small) language model in the browser