How come I get two different results when calling the embeddings API in parallel using the `examples/api_request_parallel_processor.py` script?

[luofuli]

trafficstars

I've noticed that running the examples/api_request_parallel_processor.py script twice with the same input file produces non-identical results. Only approximately half of the text embeddings show slight differences in their decimal places, with errors ranging from 1e-4 to 1e-8 in each dimension. What could be causing these discrepancies? Could it be due to bugs in the examples/api_request_parallel_processor.py script, parallel request sending, or the embedding model's inherent randomness, such as dropout?

PS: The embedding model is text-embedding-ada-002.

[luofuli]

Additionally, I would like to confirm the following questions: For a specified model (e.g., text-embedding-ada-002):

1. Does the embedding for the same text retrieved through the API change over time?
2. Does the meaning of each dimension in the embedding retrieved at different times?

[ted-at-openai]

Inherent randomness sounds like the most likely culprit. Those sound like small differences.

Nothing is intended to change over time.