Handling of missing values (NaN)

[otsaw]

Hi! Is Voyager intended to handle missing values in some way? It might be useful to allow missing values in the query vector, in which case I think the search should consider other elements only. Would that be possible given the data structure you have, even if those queries would be less efficient?

I'm trying out the Python package and if I set an element in the query vector to numpy.nan, it seems to propagate so that all distances are NaN as well and the results don't seem meaningful.

[otsaw]

For reference, what I'm looking for seems to be called "NaN-Euclidean" metric by some others:

scikit-learn/scikit-learn#25319

facebookresearch/faiss#3355

[stephen29xie]

Hello. Voyager doesn't handle missing values and we don't have any plans to at the moment. I agree, Voyager returning NaN distances when missing values are present isn't very useful. Perhaps we can throw an error when missing values are present to be clearer about the nonsupport.