Interleaving

Definition

Interleaving merges two ranked result lists into a single ranking by alternating (“zip-merge”) items from each, optionally weighted by sampling probabilities so one source contributes more often. It serves two purposes:

  1. Fusion baseline — a simple, hard-to-beat way to combine multiple retrievers (e.g. lexical + vector) when their scores are non-comparable, without training a model.
  2. Online evaluation — interleaving two rankers and observing which side’s items get clicked is a sensitive B testing alternative that needs less traffic than split testing.

Why it matters

In a multi-retriever / Hybrid Search setup, interleaving solves the cold-start problem: you need behavioral history before you can train an LTR fusion model, but you need some unified ranking to collect clicks. Interleaving provides that initial ranking, after which a LambdaMART / LTR re-ranker (e.g. Metarank) can take over.

Articles