Really nice implementation! And it's so cool to be able to do this offline. The embeddings aren't quite there yet. One trick that might be helpful is to embe

Really nice implementation! And it's so cool to be able to do this offline. The embeddings aren't quite there yet. One trick that might be helpful is to embed only the defining (usually the first) sentence or paragraph of the Wikipedia article, rather than the whole document -- not clear to me which portion you're using now. My own site, OneLook, has had a similar feature (https://onelook.com/thesaurus/ https://onelook.com/thesaurus/) since '03 that lets you find words and concepts by description. It was a pure reverse-dictionary search back when I started, but over the past two decades I've explored word embeddings, then sentence embeddings, and more recently LLMs. Nowadays it uses GPT to generate some guesses for inputs that it can't answer itself. LLMs are so much better than earlier methods at this task, it's taken some of the wind out of my sails on improving this aspect of OneLook. I frequently hear from people for whom reverse-definition lookups are the main reason they use ChatGPT!