I average the embeddings of every 512 bytes of the page text
Ah ok, that makes sense
That might actually be making things worse for longer articles. It probably would be better would be to index them separately and aggregate back to article level post query