If it’s just search, might be able to train a large BERT model which should be way cheaper than a LLM. Would need to plot out how much data you’re getting each time period (like daily) and then see how much it would cost to train BERT on that