Cosine similarity
Also written: dot product
A score for how close two embedding vectors are in meaning: 1 for the same meaning, near 0 for unrelated. For normalized vectors it equals the dot product, which is why a single matrix multiplication can score every chunk.