E.g. normalize vectors to unit norm and use dot product instead of cosine similarity as metric. Should be very fast. Some benchmarks and snippets thanks to Tim Scarfe:    