ML//neural network//embedding//cosine similarity

Measures angular alignment between two vectors: cos(θ) = (A·B)/(|A||B|). Range [-1, 1]. 1 = same direction, 0 = orthogonal, -1 = opposite.

Measures angular alignment between two vectors: cos(θ) = (A·B)/(|A||B|). Range [-1, 1]. 1 = same direction, 0 = orthogonal, -1 = opposite.

The standard metric for semantic search in RAG: encode query and documents as vectors, find highest cosine similarity.

Why cosine and not dot product? Cosine normalizes for magnitude — a long document vector shouldn't score higher just because it has more words.

After contrastive learning fine-tuning: "Capital of France?" and "Paris is the capital" → high cosine similarity.

The backbone of vector databases and sentence transformer search pipelines.