Why use cosine over Euclidean distance?

Cosine measures angle (direction), ignoring magnitude. Two documents of different lengths can still be similar if they discuss the same topics.

What is a good similarity threshold?

Depends on use case. For semantic search, 0.7+ is often considered similar. For duplicate detection, 0.95+ is common.

Can I use cosine similarity for sparse vectors?

Yes, it works well for sparse vectors like TF-IDF. Zero components don't contribute to similarity.

When to use Euclidean vs Cosine?

Euclidean considers magnitude (length) while Cosine only considers direction. Use Euclidean when magnitude matters (e.g., physical distance), Cosine when it doesn't (e.g., document similarity).

Does Euclidean distance work in high dimensions?

It suffers from the "curse of dimensionality" - in very high dimensions, all points become roughly equidistant. Consider dimensionality reduction first.

Should I normalize vectors first?

If you normalize vectors to unit length, Euclidean distance becomes related to Cosine similarity: d² = 2(1 - cos(θ)). Normalization helps when scales differ.