implementing locally run recommendation algorithm using vector db / semantic similarity:
generate vector db from data of interest. see: https://sbert.net
fetch list of liked stuff from service. then query vector db of large collection of data with this data of likes.
ie.
prompt: video titles of liked youtube videos
->
query db of large collection of data with video_ids & titles
prompt: contents of liked nostr notes
query db of large collection of nostr notes
mid grade nvidia card can easily do these queries for example against dataset of 12 M youtube video titles. encoding data of 12 M video titles to vector db takes maybe ~18h of gpu time.