We found embedding indexing bottleneck in the least expected place: JSON parsing