ValueError
builtins.ValueError
Stack trace
ValueError: cosine similarity resulted in NaN due to zero norm vector
File "embedding_utils.py", line 42, in compute_cosine_similarity
similarity = np.dot(vec1, vec2) / (np.linalg.norm(vec1) * np.linalg.norm(vec2))
File "embedding_utils.py", line 58, in main
sim = compute_cosine_similarity(embedding1, embedding2) # triggers error Why it happens
Cosine similarity divides the dot product of two vectors by the product of their norms. If either vector is a zero vector (all elements zero), its norm is zero, causing division by zero and resulting in NaN. This often happens when embeddings are empty, uninitialized, or contain invalid values.
Detection
Check for zero vectors by verifying if the norm of any embedding vector is zero before computing cosine similarity, and log or raise a clear error if detected.
Causes & fixes
One or both embedding vectors are zero vectors with all elements zero
Add a check to ensure embedding vectors have non-zero norm before computing cosine similarity; handle zero vectors by skipping or assigning similarity=0.
Embedding vectors contain NaN or infinite values due to upstream errors or corrupted data
Validate embeddings for NaN or infinite values using numpy functions like np.isnan() or np.isfinite() and clean or regenerate embeddings before similarity calculation.
Incorrect embedding extraction resulting in empty or zero-length vectors
Verify embedding extraction logic to ensure vectors are correctly generated and have expected dimensionality and values.
Code: broken vs fixed
import numpy as np
def compute_cosine_similarity(vec1, vec2):
# This line causes NaN if vec1 or vec2 is zero vector
return np.dot(vec1, vec2) / (np.linalg.norm(vec1) * np.linalg.norm(vec2))
embedding1 = np.array([0, 0, 0])
embedding2 = np.array([0.1, 0.2, 0.3])
similarity = compute_cosine_similarity(embedding1, embedding2) # triggers ValueError or NaN
print(f"Similarity: {similarity}")import numpy as np
def compute_cosine_similarity(vec1, vec2):
norm1 = np.linalg.norm(vec1)
norm2 = np.linalg.norm(vec2)
if norm1 == 0 or norm2 == 0:
return 0.0 # Avoid division by zero, define similarity as 0
return np.dot(vec1, vec2) / (norm1 * norm2)
embedding1 = np.array([0, 0, 0])
embedding2 = np.array([0.1, 0.2, 0.3])
similarity = compute_cosine_similarity(embedding1, embedding2) # fixed to avoid NaN
print(f"Similarity: {similarity}")Workaround
Wrap cosine similarity calculation in try/except to catch ValueError or check for NaN results and fallback to zero similarity or skip calculation.
Prevention
Validate all embedding vectors for zero norm, NaN, or infinite values immediately after generation and before similarity computations to ensure data integrity.