ContextQMD
Back to Fabric

Azure OpenAI Troubleshooting

docs/Azure-OpenAI-Troubleshooting.md

1.4.4523.9 KB
Original Source

Azure OpenAI Troubleshooting

This document describes a known issue with Azure OpenAI integration and its fix.

Issue: DeploymentNotFound Error (404)

Symptoms

When using Fabric with Azure OpenAI, you may encounter this error:

POST "https://{resource}.cognitiveservices.azure.com/openai/chat/completions?api-version=...": 404 Not Found
{
  "code": "DeploymentNotFound",
  "message": "The API deployment for this resource does not exist..."
}

Root Cause

Azure OpenAI requires deployment names in the URL path:

✅ Correct:  /openai/deployments/{deployment-name}/chat/completions
❌ Incorrect: /openai/chat/completions

The OpenAI Go SDK's azure.WithEndpoint() middleware has a bug in its URL transformation logic:

  1. The SDK's jsonRoutes map expects paths like /openai/chat/completions
  2. But the SDK actually sends paths like /chat/completions (without the /openai/ prefix)
  3. The /openai/ prefix is included in the base URL, not the request path
  4. This causes the route matching to always fail, so deployment names are never injected into the URL

Technical Details

In the SDK's azure/azure.go:

go
// SDK checks for these routes:
var jsonRoutes = map[string]bool{
    "/openai/chat/completions": true,  // Expects /openai/ prefix
    // ...
}

// But actual request path is:
path := "chat/completions"  // No /openai/ prefix!

The mismatch means jsonRoutes[req.URL.Path] never matches, and the deployment name transformation never happens.

Fix

The fix in internal/plugins/ai/azure/azure.go adds custom middleware that:

  1. Intercepts outgoing requests
  2. Extracts the deployment name from the request body's model field
  3. Transforms the URL path to include /deployments/{name}/
go
// Transform: /chat/completions -> /openai/deployments/{name}/chat/completions
func azureDeploymentMiddleware(req *http.Request, next option.MiddlewareNext) (*http.Response, error) {
    // Routes that need deployment name injection
    deploymentRoutes := map[string]bool{
        "/chat/completions":     true,
        "/completions":          true,
        "/embeddings":           true,
        "/audio/speech":         true,
        "/audio/transcriptions": true,
        "/audio/translations":   true,
        "/images/generations":   true,
    }

    // Extract deployment from body and transform URL...
}

Additional Fix: StreamOptions Error

Symptom

400 Bad Request
{
  "message": "The 'stream_options' parameter is only allowed when 'stream' is enabled."
}

Cause

The Chat Completions API was sending stream_options for all requests, but Azure only accepts this parameter when stream: true is also set.

Fix

Moved StreamOptions to only be set for streaming requests in internal/plugins/ai/openai/chat_completions.go.

Configuration

Ensure your Azure OpenAI configuration is correct:

bash
# In ~/.config/fabric/.env
AZURE_API_KEY=your-api-key
AZURE_API_BASE_URL=https://{your-resource}.cognitiveservices.azure.com/
AZURE_DEPLOYMENTS=your-deployment-1,your-deployment-2  # Comma-separated deployment names
AZURE_API_VERSION=2024-12-01-preview  # Optional, defaults to 2024-05-01-preview

Note: The deployment name is what you specified when deploying a model in Azure AI Foundry (formerly Azure OpenAI Studio), not the model name itself (e.g., my-gpt4-deployment rather than gpt-4).

Verification

Test your Azure OpenAI setup:

bash
fabric --model <your-deployment-name> --pattern summarize "Hello world"

Replace <your-deployment-name> with the actual deployment name from your Azure configuration.

You should see a successful response from your Azure OpenAI deployment.

References