Discussions

Hi. I have a few million documents that I generated embeddings for. Based on ad hoc experiments, it seems like for my purpose, a lower embedding dimensionality would be fine. I want to take my existing float/2048 dimension embeddings and reduce to int/1024. Obviously I can do whatever I want with my local embeddings, but is there a means to exactly replicate the quantization/dimension reduction done with the API?

In other words, is there a function available to take a float/2048 dimension embedding and change it to an int/1024 dimension embedding that exactly matches the int/1024 dimension embedding I would have gotten from the model directly, without paying to embed it again?

Thanks