Thanks for your comments :)

I've used tensor-based search before using Faiss Index and finetuned bert models (it's still in the code). As I mentioned in my article, that slowed down a bit the process since, each time a field is chosen, the bert model is loaded and took about 4 seconds more. That's why I switched to TF IDF. But I plan to optimize the tensor-search part more (I'll check Marqot!), hopefully with the help of the open source community :)