Textbook of AI

Abstract. Introduces Sentence-BERT (SBERT), the standard architecture for fixed-length sentence embeddings. Plain BERT produces token-level contextual embeddings; pooling them naively yields poor sentence similarity. SBERT fine-tunes BERT in a Siamese architecture with a contrastive objective on natural-language-inference pairs, so that semantically similar sentences map to embeddings with high cosine similarity. SBERT cut sentence-similarity inference time from $\mathcal{O}(n^2)$ pairwise BERT calls to $\mathcal{O}(n)$ embeddings plus dot-product search. Its descendants, MPNet, E5, BGE, GTE, power the modern retrieval-augmented-generation stack.