You already know: tries and prefix structures (D2), learning to rank and position bias (D7 and the e-commerce case), caching and in-memory serving (D16), and from the friend suggestions case that what a product suggests can disclose something. This case adds the tightest latency loop in search: a request on every keystroke, a few milliseconds of server time, suggestions that must reflect breaking news within minutes, and every suggestion is something the product says. The design questions: how to serve in milliseconds, where suggestions come from, how to rank them, and how to keep them safe and private.
How to read this lesson. No code, on purpose. Numbers marked assumed are assumptions you state out loud.
1. The prompt, and the questions that change the design
A mid-level answer: "a trie of past queries, return the most popular completions". A staff candidate first asks the latency budget per keystroke, where suggestions come from (past queries, entities, generated text), how fresh and personal they should be, and what safety and privacy rules apply. Those answers produce precomputed suggestions per prefix held in memory, a popular-query index plus a trending layer, light personalisation, and safety filters plus a distinct-user threshold.
Assumptions you state: