Kuba Rogut

Kuba Rogut is a deployed engineer at Turbopuffer, a vector database company, and a practitioner-advocate for agentic retrieval architectures. He is the speaker behind the talk "RAG is dead, right??" in which he argues that RAG has evolved far beyond simple vector search and remains very much alive.

Core argument: RAG is not dead, it has matured

Rogut directly counters the "RAG is dead" narrative circulating on social media, citing empirical evidence: Google search volume for RAG hit a new inflection point in mid-2025 and went through the roof, contradicting claims of its decline. 1:53

His thesis is that the framing of RAG as dead conflates the original simple pattern with the broader category. He argues RAG encompasses not just vector search but also full text search, grepping, globbing, regex, and filters — used iteratively by an agent. 3:20 What has died, in his view, is naive single-shot vector search: "a lot of people are no longer doing the simple rag" that worked in 2023 and early 2024, because sophisticated users have moved to agentic search for real performance gains. 14:57

Agentic Search as the new default

Rogut advocates strongly for Agentic Search, defining it as giving agents a set of tools to progressively and iteratively find and reason over context — explicitly distinguishing it from merely grepping through the file system. 4:11 He frames hybrid, tool-rich retrieval as "becoming a default for serious agentic search." 0:43

Embeddings as cached compute

A notable framing Rogut offers is that Embeddings and semantic search are a form of cached compute, with the value of that cache depending on query volume. 10:41 This reframes the build-vs-skip question for semantic indexing as an economic one rather than a purely technical one.

Staged retrieval vs. large context windows

Rogut pushes back on the idea that expanding context windows make retrieval obsolete. His position is that you don't need a trillion tokens at once — you need the right million, making staged retrieval essential even as context windows grow. 16:59

Cursor vs. Claude Code as a case study

Rogut uses the contrast between Cursor and Claude Code as a concrete illustration of retrieval philosophy in production. He argues that Cursor's upfront indexing approach saves tokens, time, and money compared to Claude Code's per-session grep-based discovery. 13:19 He supports this with Cursor's own reported metrics: adding semantic search produced a 24% increase in answer accuracy on their composer model and a 2.6% gain in code retention in large codebases. 8:36