Retrieval

Going and fetching the right pages before answering a word. A customer asks what your cancellation terms are; retrieval is the step that walks to the shelf, finds the two paragraphs that actually cover it, and puts them on the desk. The model (the AI itself — the part that reads and writes) only ever sees what retrieval hands it.

Three words get used for one shelf of ideas, so here is the chain: RAG is the whole open-book exam, retrieval is fetching the page, and semantic search is how you find that page when it's filed under words nobody would think to type.

Why it matters: garbage retrieved, garbage generated. When a "chat with your documents" tool gives a bad answer, the fault is nearly always here — it was handed the wrong pages and answered them faithfully. So the first question is "what was it actually shown?", not "why is the AI stupid?"

Related: RAG · Semantic search