use case · basics
The atom of everything embeddings do. Give the model two sentences; it turns each into a 384-number vector and reports the cosine similarity between them — 1.00 means "same direction in meaning space", near 0 means "unrelated". Same idea in different words scores high; different topics score low.
Try a preset pair:
Cosine similarity ranges from -1 to 1, but for this model most real sentence pairs land between about 0.1 (unrelated) and 0.9 (near-paraphrase). Notice that "masterpiece" vs "terrible" still scores fairly high — they share a topic (a film) even though the opinions flip. Embeddings capture aboutness, not agreement.
← Back to MiniLM · Next: Practical →