Cosine Similarity

When comparing embedding vectors, the most common measure is cosine similarity. It looks at the angle between two vectors — not their length — capturing whether they point in the same direction (similar meaning) regardless of how "long" they are. This makes it the default choice for text similarity and semantic search.

💡 In one line: Cosine similarity measures the angle between two vectors — the smaller the angle, the more similar the meaning.

What is Cosine Similarity?

It measures the cosine of the angle between two vectors. The value ranges from −1 to 1 (for many text embeddings it lands roughly in 0 to 1). Crucially, it depends only on direction, not magnitude — two vectors pointing the same way score high even if one is much longer.

The Formula

cosine(A,B)=A⋅B∥A∥ ∥B∥\text{cosine}(A, B) = \frac{A \cdot B}{\lVert A \rVert \, \lVert B \rVert}
  • A · B — the dot product of the two vectors.
  • ‖A‖, ‖B‖ — their magnitudes (lengths).

The result is the cosine of the angle between them.

Interpreting the Score

CosineAngleMeaning
10°Same direction — very similar
090°Perpendicular — unrelated
−1180°Opposite — opposite meaning

Higher = more similar.

Why It's Preferred for Embeddings

  • Ignores magnitude — a short and a long text about the same topic still score high.
  • Focuses on direction — which is where the meaning lives.
  • Robust in high dimensions — reliable even with hundreds of dimensions.

It's the default measure in most semantic-search and vector-database setups.

Cosine Similarity vs. Cosine Distance

They're two sides of the same coin:

cosine distance = 1 − cosine similarity
  • Similarity: higher = more alike.
  • Distance: 0 = identical, higher = more different.

Example

Illustrative scores:

  • "cat" vs "kitten" → 0.85 (very similar)
  • "cat" vs "car" → 0.15 (barely related)

Code Example


 
Tip: if you normalise vectors to length 1 first, cosine similarity becomes just the dot product — which is faster at scale.
 
When to Use It
 
     
  • Default for text and semantic similarity, where length shouldn't matter.
    •  
  • If magnitude genuinely matters, consider Euclidean distance (next subtopic).
    •  
 
Summary
 
     
  • Cosine similarity measures the angle between two vectors, ignoring magnitude.
    •  
  • Formula: dot product ÷ product of magnitudes.
    •  
  • 1 = same direction (similar), 0 = perpendicular (unrelated), −1 = opposite.
    •  
  • It's the default for embeddings because meaning lives in direction, not length.
    •  
  • Cosine distance = 1 − similarity; normalise vectors to reduce it to a fast dot product. EOF echo created