What is Cosine Similarity

How It Works

Cosine similarity measures the cosine of the angle between two vectors, providing a metric for similarity based on direction rather than magnitude. It's commonly used in vector space models to compare document similarity.

Technical Details

Calculated as the dot product of two vectors divided by the product of their magnitudes. Values range from -1 (completely dissimilar) to 1 (completely similar), with 0 indicating orthogonality.

Cosine vs Dot Product vs Euclidean (L2) vs Manhattan (L1)

• Cosine similarity: dot(a, b) / (|a| × |b|); compares direction only, from -1 to 1
• Dot product: Σ a_i × b_i; direction and length together, so longer vectors score higher
• Euclidean (L2) distance: √Σ (a_i - b_i)²; straight-line distance, lower is closer
• Manhattan (L1) distance: Σ |a_i - b_i|; sum of per-dimension differences, less sensitive to one large difference
• Example with a = (1, 0, 1) and b = (0, 1, 1): dot = 1, cosine = 1 / (1.414 × 1.414) = 0.5, L2 = 1.414, L1 = 2
• For vectors normalized to length 1, cosine, dot product and L2 rank results in the same order, since L2² = 2 - 2 × cosine
• Use the metric the embedding model was trained with; most text and image embedding models specify cosine or dot product

Distance Metrics in Mixpeek

Each vector index in a Mixpeek collection has one distance metric. Built-in extractors set it for you; a custom extractor declares it as `distance_metric` in its manifest, with `"cosine"`, `"euclid"` or `"dot"`. Manhattan (L1) is not one of the options.

Best Practices

• Normalize vectors before computing similarity
• Use for high-dimensional data comparisons
• Combine with other metrics for comprehensive analysis
• Optimize vector storage for performance
• Regularly assess similarity thresholds

Common Pitfalls

• Ignoring vector normalization
• Over-relying on cosine similarity alone
• Inadequate threshold tuning
• Poor vector storage optimization
• Lack of comprehensive similarity analysis

Advanced Tips

• Implement hybrid similarity measures
• Use dimensionality reduction for efficiency
• Consider domain-specific similarity adjustments
• Optimize for specific use cases
• Regularly review similarity performance