Similarity score calculations may be incorrect for some vector distance algorithms

[anders-swanson]

In some vector stores, I'm seeing Document.score calculated as 1-distance. This works when distance is bounded [0,1] and a lower distance value means "more similar". However, this isn't very meaningful for vector distance calculations that aren't using normalized vectors, like Euclidean Squared or Manhattan.

Are people commonly using distance metrics aside from Cosine Distance? It seems like the current implementations might not work that well otherwise.