Pairwise Cosine Similarity in NumPy
Problem statement
A retrieval service stores document embeddings as rows of a matrix and needs to score a batch of query embeddings against all of them. Implement cosine similarity in NumPy for that setting.
A is an $m \times d$ nested list (one query vector per row) and B is an $n \times d$ nested list (one document vector per row). Return the $m \times n$ matrix $S$ with
$$S_{ij} = \frac{a_i \cdot b_j}{\lVert a_i \rVert_2 \, \lVert b_j \rVert_2}$$
If either vector has zero norm, define $S_{ij} = 0.0$; the output must never contain NaN or inf. Use vectorized NumPy rather than a Python loop over pairs. Return plain nested lists (a NumPy array is converted for you). Results are compared with a relative tolerance of $10^{-6}$, so you do not need to clip to $[-1, 1]$.
def cosine_similarity(A: list, B: list) -> list:
...
Constraints
- $1 \le m, n \le 200$
- $1 \le d \le 256$
- every entry is in $[-100, 100]$
Example 1
A = [[1, 0], [1, 1]]
B = [[1, 0], [0, 1], [-1, -1]]
-> [[1.0, 0.0, -0.7071], [0.7071, 0.7071, -1.0]] (rounded for display)
The second query points at 45 degrees, so it is equally similar to both axes and exactly opposite to $(-1, -1)$.
Example 2
A = [[0, 0, 0]]
B = [[1, 2, 3]]
-> [[0.0]]
A zero vector has no direction, so its similarity is defined as 0.0.
Example 3
A = [[3, 4]]
B = [[6, 8], [-4, 3]]
-> [[1.0, 0.0]]
$(6, 8)$ is a scaled copy of $(3, 4)$; $(-4, 3)$ is orthogonal to it.
Show hint
Can every pairwise similarity be read off one matrix product if the vectors are prepared the right way first?
Relevant roles: ML Engineer, Forward Deployed Engineer and AI Engineer.
Related free problems
- Multi-Head AttentionMedium
- Masked Self-AttentionMedium