Reranker models
A reranker’s job is to provide the final, definitive judgment. While standard rerankers are generic, the K² reranker is a specialist. It examines the relationship between a query and each candidate using a deep cross-encoder, trained on the specific ambiguities and hard negatives surfaced by our retrieval stage. It doesn’t just reorder a list, it separates almost right from truly correct.
The reranker trains on the same adversarial negatives as your dense and sparse encoders, so it understands why each candidate surfaced and how to correct it.
Outputs are calibrated relevance scores, not just ranks, so you can enforce answer/deferral thresholds with confidence.
Dashboards and hooks ship with the model so you can monitor health, spot degradation, and retrain before incidents.
The final arbiter is tuned to your domain, not a generic benchmark.
Generic rerankers don't know why a document was retrieved. Ours does. By training on the same hard negatives as the retrievers, it understands their failure modes and corrects them with confidence.
Cross-encoder architectures read the query and document together (`[CLS] query [SEP] document [SEP]`) so subtle differences between almost-right and perfect answers are captured.
Outputs are calibrated relevance scores, not just a new order. Set thresholds ("only answer if score > 0.9"), prevent hallucinations, and know when to defer.