You’re in a Senior AI Engineer interview at Google and the interviewer asks:
“Your RAG system pulls from a SQL database, a private company corpus, and web search. The answers now contradict each other. How do you fix source selection?”
Don’t say: “I’d rank all the chunks by relevance and take the top-k.”
Wrong. You just told them you’d let a 2021 Reddit thread outvote your production database.
The reality: relevance is not authority. And reciprocal rank fusion treats every source as a peer.
The moment you blend three heterogeneous stores into one similarity ranking, you’re averaging witnesses instead of asking which one was actually in the room.
Here’s what source-selection accuracy actually requires: