The trade-off between speed and quality is everything though. I've found that semantic searching everything that might be relevant and dump the top N into the model is way faster at getting a workable result.