FlashRank cross-encoder reranking for frankensearch.

Provides the [FlashRankReranker] implementation of the [Reranker] trait, using ONNX Runtime for cross-encoder scoring with sigmoid activation on raw logits. Gracefully falls back to original scores if the model is unavailable.

Model Layout

Required files in the model directory:

• onnx/model.onnx (preferred) OR model.onnx (legacy)
• tokenizer.json

Architecture

Cross-encoders differ from bi-encoders: instead of comparing pre-computed embeddings, they process the query and document together through a transformer, producing direct token-level attention between them. This is dramatically more accurate but cannot pre-compute anything.

(query, document) → tokenize → ONNX → logit → sigmoid → score ∈ [0, 1]