rag

Why the Same Query Returns Different Answers: Randomness in AI Search

There is a comforting myth about AI search. It says that somewhere underneath it all sits a ranking, and if you play it right, you climb. There is no ranking. Between the question you ask and the answer you get sit six distinct stages: parsing, fan-out, retrieval, assembly, sampling, and citation. Each one introduces randomness on purpose, not by accident, and the variance compounds as it moves down the pipeline. This is a walk through all six, traced end to end through a single real query, so you can see exactly where the randomness enters, and why the whole idea of holding a fixed position falls apart.