Retrieval Pipeline
Also called RAG Pipeline
The full assembly line from a user question to the passages handed to the model.
Think of it like
A kitchen line: prep the order, pull ingredients, plate it — each station feeding the next.
Example
Query rewrite → embed → vector search → rerank → assemble context — a typical pipeline behind one chatbot reply.
How it actually works
Real retrieval is rarely one call; it is a sequence of stages — rewrite, expand, embed, search, filter, rerank, assemble — each of which can help or hurt. Treating it as a pipeline lets you measure and tune each stage independently, which is how teams debug "the answers are bad" into "reranking is dropping the right chunk". More stages add latency, so every one should earn its place.
For product teams
Seeing retrieval as stages turns vague quality complaints into fixable, measurable steps.
For engineers
Composable stages (rewrite→embed→search→filter→rerank→assemble); each is independently evaluable and a latency cost.
Related
- Retrieval — Wraps around the core retrieval step.
- Query Rewriting — Query rewriting often sits at the front.
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