PS Possibly you could add a floor (say less than 20 to exclude shill or throwaway accounts), or ignoring karma difference over a threshold (say 1000). I think a

PS Possibly you could add a floor (say less than 20 to exclude shill or throwaway accounts), or ignoring karma difference over a threshold (say 1000). I think applying simple filters based on common sense or domain knowledge can help vector similarity searches with sparse data or pollution. Just a random thought :)