I’ve got to wonder if that file size isn’t a bug. Sparse database gone nuts.
math.log(2_116_946_514, 2)
30.979337673184506
2GB is in the “mail me a flash drive” territory. Is that even in the range of possibility, that you were billed for perhaps 500M of tokens if it were language? One embeddings is at most 12k, and the cutoff should be 50000 embeddings per batch, across all the calls and their lists.