Why do people keep saying RAG is dead?

Three waves of announcements fed the obituary, and each one killed a real weakness of early RAG without killing retrieval.

Can a million-token context window replace retrieval?

Sometimes, and you should know exactly when, because the math is unforgiving at scale.

What does the research say about long-context degradation?

The uncomfortable finding of 2023 to 2025 research is that a token in the window is not a token in the model's attention.

How did RAG practice change between 2024 and 2026?

The stack that survives in production today barely resembles the 2023 tutorials I still see copied in client codebases.

When do you genuinely not need RAG?

More often than RAG vendors admit. Skip the pipeline when:

Is RAG dead in 2026? No, it just became the default plumbing

Is RAG dead in 2026? No, it got boring. What long context really changes, what context-rot studies show, and when retrieval still wins.

"RAG is dead" has been declared roughly every six months since 2023, and every time the funeral gets scheduled, retrieval quietly ships in another platform. So here is my direct answer: no, RAG is not dead in 2026. What died is the 2023 demo stack, the naive pipeline of chop-everything-into-512-token-chunks, embed, cosine-similarity, pray. What replaced it is not "no retrieval". It is retrieval so standard that nobody writes a launch post about it anymore.

The confusion is understandable. Context windows exploded, agents learned to search on their own, and every major platform now sells managed retrieval as a checkbox. If your mental model of RAG is "a vector database I have to babysit", then yes, that job is disappearing. If your mental model is "give the model the right information at the right moment, under cost, freshness and permission constraints", that job got bigger.