Retrieving less can make RAG answers more relevant
Abstracts
Abstracts
A RAG system can return a worse answer when it retrieves more context than the question needs.
In “AI Routing: The Missing Quality Layer Above RAG and MCPs,” I describe a thin routing step that classifies a query and selects the minimal data collections to search. The point is not to add another source of information; it is to keep unrelated material out of the answer context.
That distinction matters when building practical AI systems: retrieval should be judged by whether it brings in the right information, not by how much it brings in. The article argues that narrower selection can improve relevance while reducing token use.