Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks

The workload-shape paper for batch encoding: its abstract quantifies the cost of NOT having a bi-encoder - finding the most similar pair in a collection of 10,000 sentences 'requires about 50 million inference computations (~65 hours) with BERT' - i.e. 10,000 x 10,000 pairwise cross-encoder forwards. This is the primary-source statement that cross-encoder scoring is quadratic in corpus size and cannot reuse a KV cache, and it is the motivating arithmetic for the reranking flop record.