Divide your data into smaller chunks, then use some kind of initial vector similarity check to choose only the relevant bits.