Cohere Embed v4 returns empty chat.completion response - model capability potentially misconfigured - failing since January 16, 2026

[syyhavtr]

Summary

Cohere Embed v4 model on Azure AI Foundry returns HTTP 200 with an empty chat.completion response instead of embeddings. The model capability appears to be misconfigured with embeddings: false.

Environment

• Model: Cohere Embed v4 (embed-v-4-0)
• API Version: 2024-02-15-preview
• SDKs tested: OpenAI SDK, Azure AI Inference SDK, Cohere SDK

Problem

The embeddings endpoint returns the wrong response type:

Request:

POST /openai/deployments/embed-v-4-0/embeddings?api-version=2024-02-15-preview Content-Type: application/json
{"input": ["Test"], "input_type": "document"}

Actual Response (WRONG):

{"id":"","model":"","choices":[],"usage":null,"created":-1,"object":"chat.completion"}

Expected Response:

{"object":"list","data":[{"embedding":[0.1,0.2,...],"index":0}],"model":"embed-v-4-0"}

Potential Root Cause

The model capabilities appear misconfigured. Querying GET /openai/models shows:

{
  "id": "embed-v-4-0",
  "capabilities": {
    "embeddings": false,
    "chat_completion": false
  }
}

The embeddings: false setting causes Azure to route embedding requests to the wronghandler, returning an empty chat.completion object.

Impact

• Cannot use Cohere Embed v4 for production workloads on Azure AI Foundry
• Blocking adoption of the model

Related Issues

• https://learn.microsoft.com/en-us/answers/questions/2262234/cohere-embed-v3-multilingual-
  model-was-working-for
• Cohere embedding does not work with azure ai inference Azure/azure-sdk-for-python#41050

Request

Please fix the model capability registration for embed-v-4-0 to set embeddings: true.

[artur-hellmann-slideroom]

Same here!

In my case I get an issue that stream parameter is not allowed. When allowing extra-parameters the empty chat completion appears. But it's not only that model. At least the cohere multilingual-embedding-v3-also...

[artur-hellmann-slideroom]

Yep. Very depressing...
Using cohere directly is not possible for us because we promise our customers to only work on azure servers in germany west-central...
Suddenly we got errors on search requests of users. We currently removed the search over embeddings and do only search on content directly, but all of this means that we will need to badge embed a bunch of data as soon as it works again... I hope it will work when the weeks starts, currently the traffic is lower because of weekend...

[TheRabbitHoleDev]

I don't want to be a party pooper but you might need to find another embedding solution if you really need data residency in Germany. I sadly am in the same boat as you and just went validated that, ya you can get an endpoint in your Azure AI Foundry in Germany West Central but only as Global Standart, meaning data will be processed globally.

[valentinmili]

we also have a client who wants this to be deployed in west europe, and it seems that Azure has quietly removed quota support for Cohere Embed v3 (our current model) while v4 has this bug. Hopefully they manage to fix it asap.

[jonathanwanders-rb]

I've encountered a similar issue with the flux-kontext-pro model, returning the same response as mentioned here since (at least) today, where it worked fine friday morning. I managed to solve the issue by pointing directly to the providers model within my openai endpoint, in my case <resource_name>.openai.azure.com/providers/blackforestlabs/v1/flux-kontext-pro?api-version=preview

[TheRabbitHoleDev]

Just make sure that the client understand that his data might be process outside of EU. As even with data residency we only have garantee that data stays in europe and for this model there is no data residency in eu available at the moment.

[TheRabbitHoleDev]

i tryed pointing directly to the providers model within my openai endpoint :

import requests

direct_endpoint = f"{AZURE_OPENAI_ENDPOINT.rstrip('/')}/openai/deployments/embed-v-4-0/embeddings"
api_version = "2024-10-21"
model_name = "embed-v-4-0"
    
test_text = "Hello, this is a test with extra-parameters pass-through."
    
try:
    headers = {
        "Content-Type": "application/json",
        "api-key": AZURE_OPENAI_API_KEY,
        "extra-parameters": "pass-through"  # Pass unrecognized params to model
    }
        
    payload = {
        "input": [test_text],
        "model": model_name  # Specify model in payload
    }
        
    response = requests.post(
        direct_endpoint + f"?api-version={api_version}",
        headers=headers,
        json=payload
    )
        
    print(f"\n Response Status: {response.status_code}")
    print(f"\n Full Response:")
    print(response.text)
            
except Exception as e:
    import traceback
    print(f"\n❌ Error: {e}")
    traceback.print_exc()

to only get :

 Response Status: 200

 Full Response:
{"id":"","model":"","choices":[],"usage":null,"created":-1,"object":"chat.completion","prompt_filter_results":null}

[artur-hellmann-slideroom]

FYI: Working again.

[leestott]

Thanks for confirming this has been resolved.