Design AI-Powered Search

1. Frame the Problem: Bolting Seconds onto Milliseconds

"Add AI answers to our search engine" is the Google AI Overviews, Bing, and Perplexity question. Underneath the GenAI costume, it is really an integration problem. A classic search stack already answers in 300 milliseconds at 10,000 QPS. (QPS means queries per second, so this system already handles 10,000 searches every single second.) Now it has to gain a generative layer that costs seconds and cents, without harming the thing that already works