Perplexity says grep over real checkouts beats semantic search on financial questions

Perplexity says its agents answer financial-document questions more accurately when they search sandboxed copies of real checkouts with grep and standard file tools than when they rely on broad semantic search.

On an internal benchmark of 150 financial-document questions, Perplexity reports 0.627 answer accuracy for the grep approach versus about 0.37 for semantic search. The company attributes the approach to each sandbox holding a real checkout. The results come from Perplexity's own internal benchmark and have not been independently verified.