The Score That Lied
After cross-encoder reranking, every chunk carries two scores: the original vector similarity (how close the embedding was) and the cross-encoder score (how relevant the content actually is to the query). The cross-encoder reorders the results — a chunk ranked 8th by vector similarity might be 1st by cross-encoder. The reranker got this right. What it got wrong was which score it left in relevanceScore.
relevanceScore is the field downstream consumers use for gap detection, threshold filtering, and confidence signals. After reranking, it still held the vector similarity — the score from the leg that was just overruled. A chunk promoted to the top by the cross-encoder still reported its mediocre vector similarity as its relevance. Downstream logic that filtered on relevanceScore > 0.7 would discard the cross-encoder’s best picks while keeping chunks the reranker had pushed down.
The fix is small: promote the cross-encoder score to relevanceScore and stash the original vector similarity in metadata under _vectorSimilarity. After reranking, relevanceScore reflects the signal that determined the ordering. Consumers that specifically need the vector similarity can still find it.
The subtlety is in the metadata key design. RerankingLogic already uses _crossEncoderScore to carry scores between CRAG and reranking — when both decorators are active, CRAG’s quality filter produces scores that the reranker can reuse, avoiding a second BERT forward pass. That key is a data propagation channel. The _reranked stamp is an execution guard — it prevents double-application through the blocking-to-reactive bridge.
Three keys, three purposes: _crossEncoderScore for score propagation between decorators, _vectorSimilarity for preserving the original signal, _reranked for execution guarding. Conflating any two would break the system — CRAG’s score propagation would trigger the reranker’s guard, or the preserved vector similarity would be mistaken for a pre-computed reranking score. Each key exists because the three concerns are genuinely independent.