Ettin Reranker: How a Neural Network Improves Search Accuracy

Hugging Face has released Ettin Reranker, a neural network for reranking search results. It is a family of 6 models trained by distillation from a larger model.

What is a reranker, and how does it differ from regular embedding models?

A reranker (cross-encoder) is a special class of model that takes a (query, document) pair as input and outputs a single relevance score. The key difference: a reranker encodes both sequences together, letting them see each other across all transformer layers. This yields a more accurate relevance estimate than regular embeddings.

How does a reranker differ from regular search models?

A reranker (cross-encoder) encodes the query and the document together across all transformer layers, letting them see each other. This gives a more accurate relevance score than embedding models, which encode the sequences separately.