3 ms·

I've had good success using a local model for preprocessing (tag extraction) and tuning the search parameters (hybrid search).