The LLM re-ranker evaluates candidate snippets or full pages by running multiple randomized samples through a target model (such as Gemini, Claude, or GPT) to produce a median rank, a canonical ranking, and per-item rationale for each round.
Referenced by
