Uses Cursor's measured numbers to puncture "RAG is dead" and spell out what retrieval has actually become
RAG is dead, right?? — Kuba Rogut, Turbopuffer · Kuba Rogut
11 min total·Actually worth watching closely: ~2 min·2 must-watch clips
- 0:14 – 0:55Listen
Setting up the target
He opens by stating the claim going around social media in full: throw out the vector DB, just let the agent grep through files. He repeats it back whole, as the thing this talk is going to take apart.
The argument isn't about whether to retrieve, it's about what retrieval should look like, and those two keep getting conflated.
Pure spoken setup; the slide is just a title line, so you can listen while doing something else.▶ Jump to 0:14Speaker · Kuba Rogut - 0:55 – 2:10Skim
The data isn't on the doubters' side
He pulls up Google search volume over the last couple of years and points out that demand for retrieval and search rose rather than fell, hitting a new inflection point midway through 2025. This is his first piece of evidence against "dead".
Real demand is climbing; what's being written off is a shape, not the need itself.
The screen shows a trend line. Glance at where the inflection sits and how steep the climb is; no need to follow the narration line by line.▶ Jump to 0:55Speaker · Kuba Rogut - 2:10 – 4:00Listen
The narrowed definition of RAG
He clears up a common misreading: retrieval in RAG was never just vector search. It includes full text search using something like BM25, grepping and globbing, regex, and other basic filters. Equating RAG with a single call to a vector DB is arguing with a straw man you built yourself.
What's actually outdated is the 2023 to early 2024 pattern of one vector query dumped straight into the context window, not retrieval itself.
This runs entirely on the speaker's spoken breakdown; the slide is just bullet points, so catching where each definition ends matters most.▶ Jump to 2:10Speaker · Kuba Rogut - 4:00 – 6:15Listen
What agentic search actually is
He gives the definition he thinks is accurate: not file system grep replacing a database, but giving agents a set of tools to progressively and iteratively find and reason over context. Claude Code is the example, grepping through the file system, reading a file, deciding it hasn't found what it needed, and going again until it can carry on with the task.
Agentic retrieval doesn't replace RAG; it upgrades how RAG runs, with several steps of reasoning and switching between semantic and full text as needed.
This is the conceptual hinge of the talk, carried by examples and reasoning; there's nothing you need to watch.▶ Jump to 4:00Speaker · Kuba Rogut - 6:15 – 7:17Skim
Cursor's measured gains
He shows Cursor's internal context benchmark: hooking models up to semantic search raises answer accuracy by roughly 12.5% to 13.5% on average across models, with nearly 24% for their Composer model, backed by an online A/B test.
The payoff from semantic indexing shows up in both a benchmark and a live experiment; it isn't a theoretical nice-to-have.
The screen has the comparison numbers grouped together, so reading the table beats hearing them read aloud, and there's little interpretation beyond the numbers.▶ Jump to 6:15Speaker · Kuba Rogut - 7:17 – 8:50Listen
The hidden bill for not indexing
He works through the cost of repetition: with no index, the same code base gets re-explored by different agent sessions across different developers, one sub-step alone burning thousands of tokens, and the bill compounds with every person and every day.
Make parsing and embedding a one-time upfront cost and runtime collapses to a lightweight query. What you save is tokens, time, and real money.
The cost reasoning is walked through out loud with only sparse visuals; your ears are enough.▶ Jump to 7:17Speaker · Kuba Rogut - 8:50 – 10:10Listen
Merkle trees spread the team's cost
How Cursor keeps indexing cost from scaling linearly with team size: Merkle trees compute how similar the code bases teammates open are, and if they're similar enough the existing index data is copied over, with only the files that genuinely changed rechunked and reembedded.
The "indexing is too expensive" objection can be engineered away, and this is the most directly borrowable trick in the talk.
The mechanism is mostly narrated; grasping "compare similarity, recompute only the diff" doesn't depend on the slide.▶ Jump to 8:50Speaker · Kuba Rogut - 10:10 – 10:53Listen
Bigger context still needs retrieval
He closes with Google's Jeff Dean: it doesn't matter if you get to a trillion-token context window. What you need is staged retrieval, narrowing a trillion tokens down to the right million for this particular turn.
Bigger context doesn't kill retrieval; it shifts retrieval's job from finding the answer to choosing what belongs in the window.
One quote plus a closing summary. This is where the talk lands, and hearing it closes the loop.▶ Jump to 10:10Speaker · Kuba Rogut