Cosine Similarity

Cosine similarity measures the cosine of the angle between two vectors, focusing purely on their directional alignment rather than their length. This makes it ideal for comparing text embeddings where document length should not influence semantic similarity.

• A score of 1 indicates identical direction (0° angle)
• A score of 0 indicates orthogonality (90° angle, no similarity)
• A score of -1 indicates diametrically opposite direction (180° angle)

The formula is the dot product of the vectors divided by the product of their Euclidean norms: cos(θ) = (A·B) / (||A|| × ||B||)

By dividing the dot product by the magnitudes, cosine similarity implicitly L2-normalizes both vectors to unit length. This projects all vectors onto a unit hypersphere, where only angular distance matters.

• Eliminates bias from varying document lengths in TF-IDF or embedding vectors
• Ensures a short document and a long document on the same topic can achieve a perfect score of 1
• Mathematically equivalent to the dot product of L2-normalized vectors, enabling computational optimization in vector databases

Cosine similarity is the default scoring function in most vector database and semantic search systems. When a user query is converted into an embedding, the database retrieves documents whose vectors have the smallest angular distance to the query vector.

• Used in approximate nearest neighbor (ANN) indexes like HNSW and IVF
• Enables retrieval-augmented generation (RAG) systems to find contextually relevant passages
• Pairs with TF-IDF vectorization, word2vec, and modern transformer embeddings from models like BERT and Ada

While Euclidean distance measures the straight-line distance between vector endpoints, cosine similarity measures angular separation. This distinction is critical in high-dimensional spaces.

• Euclidean distance is sensitive to magnitude; two documents with identical word distributions but different lengths will appear far apart
• Cosine similarity ignores magnitude, capturing topical similarity even when one document is a verbatim superset of the other
• In L2-normalized spaces, there is a direct mathematical relationship: cosine_sim = 1 - (euclidean_distance² / 2)

When using certain embedding models, cosine similarity can yield negative values, indicating vectors pointing in opposing directions. This property is exploited in contrastive learning and recommendation systems.

• Negative scores represent semantic opposition or dissimilarity
• In word embedding spaces, antonyms often have high cosine similarity but negative scores in specific dimensions
• Contrastive loss functions like InfoNCE explicitly push negative pairs apart in angular space, using cosine similarity as the core comparison metric

Cosine similarity is computationally efficient for high-dimensional sparse vectors, as the dot product only requires multiplication of non-zero dimensions. For dense embeddings, optimized BLAS libraries and GPU acceleration make billion-scale comparisons feasible.

• Sparse vector dot product complexity: O(nnz) where nnz is the number of non-zero entries
• Dense vector dot product: O(d) where d is dimensionality, heavily optimized in hardware
• Vector databases use product quantization and locality-sensitive hashing to avoid exhaustive cosine similarity calculations