[Bug]: `cohere-embed-v3` models fail via litellm proxy while they work fine directly via Azure AI
Closed this issue · 1 comments
taralika commented
What happened?
Make the following request via proxy:
curl https://proxy-host/embeddings \
-H 'Content-Type: application/json' \
-H "api-key: $api_key" \
-d '{"model":"cohere-embed-v3-english","input":["Dogs","Cats"]}'
response:
{
"error": {
"message": "litellm.BadRequestError: Azure_aiException - {\"error\":{\"code\":\"Invalid input\",\"message\":\"{\\\"message\\\":\\\"invalid data uri\\\"}\",\"status\":422}}\nReceived Model Group=cohere-embed-v3-english\nAvailable Model Group Fallbacks=None",
"type": null,
"param": null,
"code": "400"
}
}
Now, make the same request directly to the same model in Azure:
curl https://cohere-embed-v3-english-8834he8u0dskhj3248.westus.models.ai.azure.com/embeddings \
-H "Authorization: Bearer $azure_token" \
-H "Content-Type: application/json" \
-d '{"input":["Dogs","Cats"]}'
works fine, response:
{
"id": "324a34-6e31-43ec-a2de-e393b96faab4",
"object": "list",
"data": [
{
"index": 0,
"object": "embedding",
"embedding": [
-0.0022850037,
0.04220581,
...
]
},
{
"index": 1,
"object": "embedding",
"embedding": [
-0.0025520325,
0.024780273,
...
]
}
],
"model": "embed-english-v3.0",
"usage": {
"prompt_tokens": 2,
"completion_tokens": 0,
"total_tokens": 2
}
}
Noticed the issue with both cohere-embed-v3-english and cohere-embed-v3-multilingual models. The issue doesn't happen always, but it does happen for specific inputs like the one given in above example.
E.g., a request like this works fine via the proxy:
curl https://proxy-host/embeddings \
-H 'Content-Type: application/json' \
-H "api-key: $api_key" \
-d '{"model":"cohere-embed-v3-english","input":["Dogs are my friends","Cats are my friends"]}'
But again, the following request fails:
curl https://proxy-host/embeddings \
-H 'Content-Type: application/json' \
-H "api-key: $api_key" \
-d '{"model":"cohere-embed-v3-english","input":["Dogsaremyfriends", "Catsaremyfriends"]}'
Relevant log output
No response
Twitter / LinkedIn details
No response
krrishdholakia commented
able to repro