Sources checked
What problem does Sentence-BERT solve?
Nils Reimers and Iryna Gurevych describe the method in Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. BERT and RoBERTa were strong at sentence-pair tasks such as semantic textual similarity, but both sentences had to be fed in together.
Finding the most similar pair among 10,000 sentences required about 50 million inference computations, about 65 hours in their account. That cost makes BERT a poor index for search or clustering.
Sentence-BERT uses siamese and triplet network structures on a pretrained BERT model. The result is a sentence embedding.
Cosine similarity then replaces the pairwise BERT call. They report the same kind of search falling to about 5 seconds while keeping BERT’s accuracy on the similarity tasks they measured.
They also report that SBERT and SRoBERTa beat other sentence-embedding methods they compared on STS and transfer tasks. Those times and wins belong to the paper’s experiments.
How do you use the vectors?
Encode the question and the candidate passages separately, then rank by cosine similarity. Keep the passage identifier next to the vector so a hit can be opened and read.
Text embeddings is the general idea of a vector for a span of text, and vector search is why the nearest neighbor is not always the useful one.
In a retrieval pipeline, this ranking is the step before a RAG answer, not a substitute for reading the passage.
Why similarity is not truth
Two sentences can sit near each other because they share a topic and still disagree. A vector also will not tell you that a passage is current.
Use the embedding to choose what to read, then check the words.
Common questions
What is Sentence-BERT?
It is a BERT model modified, in the 2019 paper, to emit a sentence vector that can be compared with cosine similarity.
Is a sentence transformer the same thing?
Sentence Transformers is the code library the authors released for this method. The paper’s model is SBERT.
Later checkpoints in that library are separate releases.
Does a high similarity score mean the passage answers the question?
No. It means the vectors were close. Read the passage.
Sources & further reading
How this story was made
Written by Kristian Kostov with AI assistance and checked against the linked sources. Company performance claims are attributed to the company. Analysis reflects AiLookout’s interpretation; we have not independently tested the products discussed. Cover photography is illustrative and does not depict the specific announcement or product.
Our editorial standards





