I think a combination works quite well: first getting a small set of candidates from all the data using a lightweight model, and the using a heavy-duty model to

I think a combination works quite well: first getting a small set of candidates from all the data using a lightweight model, and the using a heavy-duty model to rerank the results and get the final candidates.