Design RAG & Retrieval

This reads like an AI problem. It is a distributed-search problem, and the round wants you to say so in the first two minutes. Anthropic's design round is known for posing a classic systems problem in model vocabulary, and retrieval is the clearest case of it. Two things separate this from a search-engine result page or a recommendations feed. The output is not a ranked shelf the user scrolls. It is a single answer Claude writes, one that has to stay faithful to the passages you fed it and cite them. The context window differs too. A feed has no such limit. Retrieval routinely returns more text than the window holds, which forces a packing decision a ranker never makes. Name that split early, then expect the scale knob to turn: ten million documents becomes a billion, fifty milliseconds becomes a million queries a second, and they watch whether the architecture bends or snaps.