Computation of dot-products is inexact

I wonder why the dot-product scores I get from search queries are significantly different than the dot-products I get by using NumPy. On 384-dimensional vectors, I have seen differences of up to 0.09. For 4-byte floats, this error is too large to be explained by accumulation of cut-off errors. So, what is the explanation?

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.