Semantic similarity puts a number on how close two pieces of text are in meaning. It does not count shared words. "Car rental" and "auto hire" share nothing on the surface yet mean almost the same thing, and the measure has to catch that.
The calculation runs on embeddings. Each text is passed through a model and turned into a vector with hundreds of dimensions. Similar meanings land near each other in that space. Taking the cosine of the angle between two vectors gives a score between zero and one, and the closer to one, the closer the meaning. With multilingual models, a Turkish sentence and its English equivalent can score highly against each other.
The applications are broad. Semantic search, duplicate and near-duplicate detection, matching support tickets to existing help articles, recommendation systems and retrieving relevant documents in a RAG pipeline all rest on this score.
A concrete case: a site has 900 blog posts and the content team does not know which ones overlap. Embedding every post and computing pairwise similarity surfaces 40 pairs scoring above 0.9. Those become candidates for merging or redirecting.
The score means nothing on its own. The useful threshold shifts with the dataset, so a few pairs get read by hand before any decision, just to see where the cutoff actually sits.
