Sentence-BERT (SBERT), published by Nils Reimers and Iryna Gurevych at EMNLP 2019, solved a practical bottleneck in using BERT for similarity tasks. Plain BERT is excellent at judging whether two sentences are related, but it has to process both sentences together, so comparing one query against thousands of candidates means thousands of expensive forward passes. SBERT modifies BERT with siamese and triplet network structures so that each sentence is mapped to a single fixed-length vector that can be compared to others with simple cosine similarity.
The speedup is dramatic. The authors report that finding the most similar pair within a collection of 10,000 sentences took roughly 65 hours with standard BERT but about 5 seconds with SBERT, while keeping accuracy comparable on semantic textual similarity benchmarks.
This work is the foundation of the widely used sentence-transformers library and, more broadly, of modern semantic search. Almost every retrieval-augmented generation pipeline today depends on turning text into comparable vectors quickly, exactly the capability SBERT made practical. For a business, SBERT is why searching a knowledge base by meaning rather than exact keywords became fast enough to run in production.