Every serious AI system now begins with retrieval. A model can only reason over what it can see, so the retriever decides what enters the context window. It turns the question into a vector, measures distance across millions of embeddings, and returns the nearest few. Keyword retrievers such as BM25 match on terms, dense retrievers match on meaning, and hybrid systems fuse both and rerank the results.

The pattern is older than the acronym. Search engines, recommendation feeds, code search, and retrieval-augmented generation all run the same loop: index the corpus, embed the query, fetch the top k. Quality is set at this stage, before a single word is generated. Ask well, fetch the right thing, and everything downstream improves. The word describes the mechanism itself.