Embedding Similarity Calculator

Cosine similarity measures the angle between two vectors, regardless of their magnitude. It returns a value between -1 and 1: 1 means identical direction (semantically very similar), 0 means orthogonal (unrelated), and -1 means opposite directions. For embeddings, scores above 0.8 typically indicate strong similarity; 0.5–0.8 is moderate; below 0.5 is weak.

Dot product measures both direction and magnitude, while cosine similarity normalizes the vectors first (making it magnitude-invariant). For normalized embedding vectors (which most models produce), dot product and cosine similarity give the same ranking. OpenAI embeddings are normalized by default, so either metric works equally for ranking.

Embeddings are high-dimensional (typically 768–3072 dimensions) numeric representations of text. Similar texts produce vectors that are close together in this high-dimensional space. They're generated by models like text-embedding-3-small (OpenAI), all-mpnet-base-v2 (Sentence Transformers), or Cohere Embed. Vector databases like Pinecone, Weaviate, and pgvector store and query these efficiently.

Use the Embeddings API: client.embeddings.create(model='text-embedding-3-small', input='your text'). The response includes a 1536-dimension float array. text-embedding-3-small is cost-effective ($0.02/1M tokens); text-embedding-3-large (3072 dims) provides higher accuracy for RAG and semantic search. Never use completion models (GPT-4o) for embeddings — use dedicated embedding models.

For RAG retrieval thresholds: >0.85 is very high confidence (near-identical content), 0.70–0.85 is relevant (good retrieval candidate), 0.50–0.70 is loosely related (use with caution), <0.50 is likely irrelevant. These thresholds vary by embedding model — newer models (text-embedding-3) tend to have tighter score distributions than older ones. Always tune thresholds empirically on your dataset.