Similarity calculation

[ceyxasm]

Dear team, I using your model and code am able to generate embeddings. However what would the preferred way to calculate a similarity score between 2 faces.
Human repo uses your model (speculatively) to generate face descriptors and from it they calculate a score using the logic here

However this results in a scenario where the scores are not super discriminative.

Here modi-2 and kamala-1 are different people but their scores are in close agreement.
Which makes me doubt the scoring logic.
Kindly help