Retrieval
The step that goes and finds the right supporting text before the model answers — the “R” in RAG.
Think of it like
Sending a runner to the archive to fetch the relevant files before the meeting starts.
Example
You ask about refund policy; retrieval grabs the refund section from the docs to hand the model.
How it actually works
Retrieval turns your question into a query, searches an index, and returns the closest matches. It can be semantic (embedding similarity), lexical (keyword/BM25), or hybrid. Everything downstream depends on it — the model can only reason over what retrieval actually fetched.
For product teams
Get this right and answers stay grounded; get it wrong and no model can save the response.
For engineers
Query-to-index matching (dense, sparse, or hybrid) returning top-k context for generation.
Related
- RAG — The engine of RAG.
- Embeddings — Searches over embeddings.
- Reranking — Often cleaned up by reranking.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome