Reranker models
A reranker provides the final relevance check before generation. The K² reranker examines each query and candidate together using a cross-encoder trained on the ambiguities and hard negatives surfaced by the retrieval stage. It does more than reorder a list: it separates stronger evidence from plausible near-misses.
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 or deferral thresholds per workflow.
Dashboards and hooks ship with the model so you can monitor health, spot degradation, and retrain before incidents.
The final ranking stage is tuned to your domain and evaluated on the queries that matter to it.
Training the reranker on the same hard negatives as the retrievers helps it learn recurring failure modes and score candidates against the same relevance objective.
Cross-encoder architectures read the query and document together (`[CLS] query [SEP] document [SEP]`) to distinguish strong matches from plausible near-misses.
Outputs are calibrated relevance scores, not just a new order. Set answer thresholds, reduce unsupported responses, and define when the system should defer.