Large-Scale Document Retrieval
A system design round. Forty-five minutes, no code. The interviewer wants to talk about search, ranking and retrieval, and how a language model sits on top of them. Expect the requirements to move.
The questions in this round
Design a retrieval-augmented assistant over ten million internal documents. Take me from a user question to a grounded answer.
One thing I left out: those documents carry per-user permissions, and two thirds of them are visible to nobody in this conversation.
And the documents change. A policy edited this morning must not be answered this afternoon from last month’s text.
Which part of this fails first at ten million documents, and how would you know before a customer tells you?