[BUG] Different embedding 3 vectors in Azure vs. OpenAI

[thai-op]

Describe the bug
When you request the same input, same api version, same model for text embedding large 3 in Azure vs. in OpenAI, you get slightly different results on the vector. They are small in floating values (< 0.0001), but in aggregate they are different enough to get bad ranking results when we mix the two together.

I don't see any documentation describing this behavior so I'm just asking here in case anyone from the Azure team knows an answer to this.

curl -H 'Authorization:Bearer <key>' -H 'Content-Type: application/json' https://api.openai.com/v1/embeddings\?api-version\=2024-03-01-preview -d "{\"input\":\"Pancreatitis in Dogs\",\"model\":\"text-embedding-3-large\"}" -o /tmp/embedding-2024-03-01-preview.openai.json

curl -H 'api-key: <key>' -H 'Content-Type: application/json' https://<region_deployment>.azure.com/openai/deployments/<deployment>/embeddings\?api-version\=2024-03-01-preview -d "{\"input\":\"Pancreatitis in Dogs\",\"model\":\"text-embedding-3-large\"}" -o /tmp/embedding-2024-03-01-preview.azure.json

diff /tmp/embedding-2024-03-01-preview.azure.json /tmp/embedding-2024-03-01-preview.openai.json -y | less

{                                                                       {
  "object": "list",                                                       "object": "list",
  "data": [                                                               "data": [
    {                                                                       {
      "object": "embedding",                                                  "object": "embedding",
      "index": 0,                                                             "index": 0,
      "embedding": [                                                          "embedding": [
        -0.01958797,                                            |               -0.019557063,
        0.00020787802,                                          |               0.00021087051,
        -0.013025163,                                           |               -0.013026878,
        0.05607405,                                             |               0.055992138,
        0.02763522,                                             |               0.027616536,
        0.012355488,                                            |               0.012345954,
        0.014163609,                                            |               0.014165475,
        -0.019387066,                                           |               -0.019389622,
        0.03151933,                                             |               0.031568136,
        0.0012730785,                                           |               0.0012551069,
        0.009816307,                                            |               0.009778531,
        0.016228437,                                            |               0.016241739,
        0.036809757,                                            |               0.03685926,
        -0.045493197,                                           |               -0.04547687,
        -0.023371628,                                           |               -0.023374708,

Exception or Stack Trace
See above

To Reproduce
See above

Code Snippet
Add the code snippet that causes the issue.

Expected behavior
The vectors should be exactly the same

Setup (please complete the following information):
n/a

Information Checklist
Kindly make sure that you have added all the following information above and checkoff the required fields otherwise we will treat the issuer as an incomplete report

• Bug Description Added
• Repro Steps Added
• Setup information Added

[alzimmermsft]

Hi @thai-op, given your sample in this issue is using cURL, and not the azure-ai-openai SDK, this seems to be a service issue and not an SDK issue, is that correct?

[thai-op]

Right, it’s not an sdk issue. The API service is the issue but I don’t know where to raise it so here I am.

…

On Fri, May 17, 2024 at 9:30 AM Alan Zimmer ***@***.***> wrote: Hi @thai-op <https://github.com/thai-op>, given your sample in this issue is using cURL, and not the azure-ai-openai SDK, this seems to be a service issue and not an SDK issue, is that correct? — Reply to this email directly, view it on GitHub <#40243 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/A7WITL4ENIXJTSPBBKPCQSDZCZLGDAVCNFSM6AAAAABH4UZE2CVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDCMJYGI2DSMZTGE> . You are receiving this because you were mentioned.Message ID: ***@***.***>

[alzimmermsft]

@brandom-msft @jpalvarezl do you know where this feedback should be re-routed to?