Definition
Plain language
Letting a model look things up in an external source and use what it finds to answer.
As stated in the literature
Retrieval-Augmented Generation, an inference pattern where a model retrieves relevant documents from an external store and conditions generation on them.
Also called: RAG poisoning
Why it matters: It lets a model stay accurate on fresh or proprietary information without retraining, and it makes answers auditable since you can see which documents were retrieved.
For example, a customer-support bot looks up the user's order in a database and the relevant help article before composing its reply.
Heard on the show
“New workload patterns — code editing where you can predict most of the output, RAG with massive shared prefixes, streaming speech.”Episode 027 — When AI Agents Build the Serving Stack: A Bet on Bespoke Infrastructure