RAG is dead, right?? — Kuba Rogut, Turbopuffer

TL;DR

• 1

  RAG is not just vector search: Rogut argues retrieval in retrieval-augmented generation includes embeddings, BM25, grep, regex, globbing, and filters, not just a single semantic lookup.

• 2

  Agentic search usually means iterative tool use: Instead of one retrieval call, agents repeatedly search, read, assess, and search again until they have enough context to act.

• 3

  Cursor gets measurable gains from semantic search: Citing Cursor's 2026 posts, Rogut says semantic search improved answer accuracy by roughly 12.5 to 13.5 percent on Cursor's internal benchmark, with one composer model showing nearly a 24 percent lift.

• 4

  Small product metrics can still matter: Cursor's online A/B test showed about 2.6 percent better retention on large codebases and a 2.2 percent drop in dissatisfied requests, even though semantic search is only helpful on a subset of queries.

• 5

  Claude Code made a different tradeoff: Rogut references Boris Cherney saying early Claude Code tried local vector DB-based RAG, then moved away from it, favoring per-session grep-style discovery instead.

• 6

  The real bottleneck is narrowing context, not expanding it: Using Jeff Dean's line, Rogut says even trillion-token windows do not remove the need for staged retrieval. "You don't need a trillion at once, you need the right million."