Do Million-Token Context Windows Make Retrieval Obsolete?

A developer notes that many recent models now ship with one-million-token context windows, prompting the question of whether retrieval pipelines are still needed. They argue that large windows are not the only consideration, pointing to cheaper cached prompt reads and retrieval failures such as missing join tables in text-to-SQL designs. The piece lists key tradeoffs: data freshness, per-token cost, latency, uneven accuracy across long contexts, and permission enforcement. It suggests caching small, stable content, retrieving large or per-user data at question time, and combining both for mixed cases. The author recommends measuring each approach against real questions before choosing.