Getting same cluster ids across multiple faiss kmeans run?

[Nitish1814]

Suppose I have vector embeddings original = [e1, e2, e3] and I run faiss kmeans over it. I will get some set of centroids.

Now, I added few more embeddings, updated = [e1, e2, e3, e4, e5] and run faiss kmeans again over it.

So my question -> is there a way such that kmeans.index.search() will give same clusterId to the same embedding e1 across multiple kmeans run ?

I have read about passing previously computed centroids as an initialiser to current train but do it works? or it is not even possible?

Thanks

[trang-nm-nguyen]

I don't believe that there is a way that there is a way to return to the same cluster to embedding across multiple Kmeans run. The option to add init centroids will initialize train with a fixed centroids, but with additional input, throughout the iterations, centroids will be recalculated, and there is no guarantee in returning a same centroid to an embedding.