Integrate with Azure OpenAI

Note

Problem-solving guide for Azure OpenAI integration

This guide helps you solve specific problems when integrating HoneyHive with Azure OpenAI, with support for multiple instrumentor options.

This guide covers Azure OpenAI integration with HoneyHive’s BYOI architecture, supporting both OpenInference and Traceloop instrumentors.

Compatibility

Problem: I need to know if my Python version and Azure OpenAI SDK version are compatible with HoneyHive.

Solution: Check the compatibility information below before installation.

Python Version Support

Support Level	Python Versions
Fully Supported	3.11, 3.12, 3.13
Not Supported	3.10 and below

Provider SDK Requirements

Minimum: openai >= 1.0.0
Recommended: openai >= 1.10.0
Tested Versions: 1.10.0, 1.11.0, 1.12.0

Instrumentor Compatibility

Instrumentor	Status	Notes
OpenInference	Fully Supported	Full Azure OpenAI support with deployment-specific tracing
Traceloop	Fully Supported	Enhanced metrics with Azure-specific cost tracking and quotas

Known Limitations

Deployment Names: Must configure Azure deployment names separately from model names
API Versions: Requires Azure API version in configuration, traced in metadata
Managed Identity: Supported but requires additional Azure SDK configuration
Streaming: Fully supported with both instrumentors

Note

For the complete compatibility matrix across all providers, see Multi-Provider Integration.

Choose Your Instrumentor

Problem: I need to choose between OpenInference and Traceloop for Azure OpenAI integration.

Solution: Choose the instrumentor that best fits your needs:

OpenInference: Open-source, lightweight, great for getting started
Traceloop: Enhanced LLM metrics, cost tracking, production optimizations

Best for: Open-source projects, simple tracing needs, getting started quickly

# Recommended: Install with Azure OpenAI integration
pip install honeyhive[openinference-azure-openai]

# Alternative: Manual installation
pip install honeyhive openinference-instrumentation-openai openai>=1.0.0

from honeyhive import HoneyHiveTracer
from openinference.instrumentation.openai import OpenAIInstrumentor
import openai
import os

# Environment variables (recommended for production)
# .env file:
# HH_API_KEY=your-honeyhive-key
# AZURE_OPENAI_API_KEY=your-azure-openai-key

# Step 1: Initialize HoneyHive tracer first (without instrumentors)
tracer = HoneyHiveTracer.init(
    project="your-project"  # Or set HH_PROJECT environment variable
)  # Uses HH_API_KEY from environment

# Step 2: Initialize instrumentor separately with tracer_provider
instrumentor = OpenAIInstrumentor()
instrumentor.instrument(tracer_provider=tracer.provider)

# Basic usage with error handling
try:
    from openai import AzureOpenAI

    # Create Azure OpenAI client
    client = AzureOpenAI(
        api_key=os.getenv("AZURE_OPENAI_API_KEY"),
        api_version="2024-02-01",
        azure_endpoint=os.getenv("AZURE_OPENAI_ENDPOINT")
    )

    # Chat completion
    response = client.chat.completions.create(
        model="gpt-35-turbo",  # Your deployment name
        messages=[{"role": "user", "content": "Hello from Azure OpenAI!"}]
    )
    # Automatically traced! ✨
except openai.APIError as e:
    print(f"Azure OpenAI API error: {e}")
except Exception as e:
    print(f"Unexpected error: {e}")

from honeyhive import HoneyHiveTracer, trace, enrich_span
from honeyhive.models import EventType
from openinference.instrumentation.openai import OpenAIInstrumentor
import openai

# Initialize with custom configuration
# Step 1: Initialize HoneyHive tracer first (without instrumentors)
tracer = HoneyHiveTracer.init(
    api_key="your-honeyhive-key",  # Or set HH_API_KEY environment variable
    project="your-project",        # Or set HH_PROJECT environment variable
    source="production"            # Or set HH_SOURCE environment variable
)

# Step 2: Initialize instrumentor separately with tracer_provider
instrumentor = OpenAIInstrumentor()
instrumentor.instrument(tracer_provider=tracer.provider)

@trace(tracer=tracer, event_type=EventType.chain)
def multi_deployment_azure_workflow(prompts: List[str]) -> dict:
    """Advanced example with business context and multiple Azure OpenAI calls."""
    from openai import AzureOpenAI

    # Configure Azure OpenAI client
    client = AzureOpenAI(
        api_key=os.getenv("AZURE_OPENAI_API_KEY"),
        api_version="2024-02-01",
        azure_endpoint=os.getenv("AZURE_OPENAI_ENDPOINT")
    )

    # Add business context to the trace
    enrich_span({
        "business.input_type": type(prompts).__name__,
        "business.use_case": "multi_deployment_analysis",
        "azure-openai.strategy": "azure_deployment_comparison",
        "instrumentor.type": "openinference"
    })

    try:
        # Test multiple Azure OpenAI deployments
    deployments = [
        "gpt-35-turbo",      # Your GPT-3.5 deployment
        "gpt-4",             # Your GPT-4 deployment
        "gpt-4-turbo"        # Your GPT-4 Turbo deployment
    ]

    results = []
    for prompt in prompts:
        deployment_results = {}

        for deployment in deployments:
            try:
                # Test each deployment
                response = client.chat.completions.create(
                    model=deployment,
                    messages=[
                        {"role": "user", "content": prompt}
                    ],
                    max_tokens=150,
                    temperature=0.7
                )

                deployment_results[deployment] = {
                    "content": response.choices[0].message.content,
                    "tokens": response.usage.total_tokens,
                    "prompt_tokens": response.usage.prompt_tokens,
                    "completion_tokens": response.usage.completion_tokens
                }

            except Exception as e:
                deployment_results[deployment] = {"error": str(e)}

        results.append({
            "prompt": prompt,
            "deployment_responses": deployment_results
        })

        # Add result metadata
        enrich_span({
            "business.successful": True,
            "azure-openai.models_used": ["gpt-35-turbo", "gpt-4", "gpt-4-turbo"],
            "business.result_confidence": "high"
        })

        return {{RETURN_VALUE}}

    except openai.APIError as e:
        enrich_span({
            "error.type": "api_error",
            "error.message": str(e),
            "instrumentor.source": "openinference"
        })
        raise

Common OpenInference Issues:

Missing Traces

# Use correct initialization pattern
# Step 1: Initialize HoneyHive tracer first (without instrumentors)
tracer = HoneyHiveTracer.init(
    project="your-project"  # Or set HH_PROJECT environment variable
)

# Step 2: Initialize instrumentor separately with tracer_provider
instrumentor = OpenAIInstrumentor()
instrumentor.instrument(tracer_provider=tracer.provider)

Performance for High Volume

# OpenInference uses efficient span processors automatically
# No additional configuration needed

Multiple Instrumentors

# You can combine OpenInference with other instrumentors
from openinference.instrumentation.openai import OpenAIInstrumentor
 from openinference.instrumentation.anthropic import AnthropicInstrumentor

 # Step 1: Initialize HoneyHive tracer first (without instrumentors)
 tracer = HoneyHiveTracer.init(
     project="your-project"  # Or set HH_PROJECT environment variable
 )

 # Step 2: Initialize instrumentors separately with tracer_provider
 # REPLACE_WITH_INSTRUMENTOR_SETUP
         OpenAIInstrumentor(),      # Works for both OpenAI and Azure OpenAI
        AnthropicInstrumentor()
    ]
)

Environment Configuration

# HoneyHive configuration
export HH_API_KEY="your-honeyhive-api-key"
export HH_SOURCE="production"

# Azure OpenAI configuration
export AZURE_OPENAI_API_KEY="your-azure-openai-key"
export AZURE_OPENAI_ENDPOINT="https://your-resource.openai.azure.com/"
export AZURE_OPENAI_API_VERSION="2024-02-01"

Best for: Production deployments, cost tracking, enhanced LLM observability

# Recommended: Install with Traceloop Azure OpenAI integration
pip install honeyhive[traceloop-azure-openai]

# Alternative: Manual installation
pip install honeyhive opentelemetry-instrumentation-openai openai>=1.0.0

from honeyhive import HoneyHiveTracer
from opentelemetry.instrumentation.openai import OpenAIInstrumentor
import openai
import os

# Environment variables (recommended for production)
# .env file:
# HH_API_KEY=your-honeyhive-key
# AZURE_OPENAI_API_KEY=your-azure-openai-key

# Step 1: Initialize HoneyHive tracer first (without instrumentors)
tracer = HoneyHiveTracer.init(
    project="your-project"  # Or set HH_PROJECT environment variable
)  # Uses HH_API_KEY from environment

# Step 2: Initialize Traceloop instrumentor separately with tracer_provider
instrumentor = OpenAIInstrumentor()
instrumentor.instrument(tracer_provider=tracer.provider)

# Basic usage with automatic tracing
try:
    from openai import AzureOpenAI

    # Create Azure OpenAI client
    client = AzureOpenAI(
        api_key=os.getenv("AZURE_OPENAI_API_KEY"),
        api_version="2024-02-01",
        azure_endpoint=os.getenv("AZURE_OPENAI_ENDPOINT")
    )

    # Chat completion
    response = client.chat.completions.create(
        model="gpt-35-turbo",  # Your deployment name
        messages=[{"role": "user", "content": "Hello from Azure OpenAI!"}]
    )
    # Automatically traced by Traceloop with enhanced metrics! ✨
except openai.APIError as e:
    print(f"Azure OpenAI API error: {e}")
except Exception as e:
    print(f"Unexpected error: {e}")

from honeyhive import HoneyHiveTracer, trace, enrich_span
from honeyhive.models import EventType
from opentelemetry.instrumentation.openai import OpenAIInstrumentor
import openai

# Initialize HoneyHive with Traceloop instrumentor
# Step 1: Initialize HoneyHive tracer first (without instrumentors)
tracer = HoneyHiveTracer.init(
    api_key="your-honeyhive-key",  # Or set HH_API_KEY environment variable
    project="your-project",        # Or set HH_PROJECT environment variable
    source="production"            # Or set HH_SOURCE environment variable
)

# Step 2: Initialize instrumentor separately with tracer_provider
instrumentor = OpenAIInstrumentor()
instrumentor.instrument(tracer_provider=tracer.provider)

@trace(tracer=tracer, event_type=EventType.chain)
def multi_deployment_azure_workflow(prompts: List[str]) -> dict:
    """Advanced example with business context and enhanced LLM metrics."""
    from openai import AzureOpenAI

    # Configure Azure OpenAI client
    client = AzureOpenAI(
        api_key=os.getenv("AZURE_OPENAI_API_KEY"),
        api_version="2024-02-01",
        azure_endpoint=os.getenv("AZURE_OPENAI_ENDPOINT")
    )

    # Add business context to the trace
    enrich_span({
        "business.input_type": type(prompts).__name__,
        "business.use_case": "multi_deployment_analysis",
        "azure-openai.strategy": "cost_optimized_azure_deployment_comparison",
        "instrumentor.type": "openllmetry",
        "observability.enhanced": True
    })

    try:
        # Test multiple Azure OpenAI deployments
    deployments = [
        "gpt-35-turbo",      # Your GPT-3.5 deployment
        "gpt-4",             # Your GPT-4 deployment
        "gpt-4-turbo"        # Your GPT-4 Turbo deployment
    ]

    results = []
    for prompt in prompts:
        deployment_results = {}

        for deployment in deployments:
            try:
                # Test each deployment
                response = client.chat.completions.create(
                    model=deployment,
                    messages=[
                        {"role": "user", "content": prompt}
                    ],
                    max_tokens=150,
                    temperature=0.7
                )

                deployment_results[deployment] = {
                    "content": response.choices[0].message.content,
                    "tokens": response.usage.total_tokens,
                    "prompt_tokens": response.usage.prompt_tokens,
                    "completion_tokens": response.usage.completion_tokens
                }

            except Exception as e:
                deployment_results[deployment] = {"error": str(e)}

        results.append({
            "prompt": prompt,
            "deployment_responses": deployment_results
        })

        # Add result metadata
        enrich_span({
            "business.successful": True,
            "azure-openai.models_used": ["gpt-35-turbo", "gpt-4", "gpt-4-turbo"],
            "business.result_confidence": "high",
            "openllmetry.cost_tracking": "enabled",
            "openllmetry.token_metrics": "captured"
        })

        return {{RETURN_VALUE}}

    except openai.APIError as e:
        enrich_span({
            "error.type": "api_error",
            "error.message": str(e),
            "instrumentor.error_handling": "openllmetry"
        })
        raise

Common Traceloop Issues:

Missing Traces

# Ensure Traceloop instrumentor is passed to tracer
from opentelemetry.instrumentation.openai import OpenAIInstrumentor

# Step 1: Initialize HoneyHive tracer first (without instrumentors)
tracer = HoneyHiveTracer.init(
    project="your-project"  # Or set HH_PROJECT environment variable
)

# Step 2: Initialize instrumentor separately with tracer_provider
instrumentor = OpenAIInstrumentor()
instrumentor.instrument(tracer_provider=tracer.provider)

Enhanced Metrics Not Showing

# Ensure you're using the latest version
# pip install --upgrade opentelemetry-instrumentation-openai

# The instrumentor automatically captures enhanced metrics
from opentelemetry.instrumentation.openai import OpenAIInstrumentor
# Step 1: Initialize HoneyHive tracer first (without instrumentors)
tracer = HoneyHiveTracer.init(
    project="your-project"  # Or set HH_PROJECT environment variable
)

# Step 2: Initialize instrumentor separately with tracer_provider
instrumentor = OpenAIInstrumentor()
instrumentor.instrument(tracer_provider=tracer.provider)

Multiple Traceloop Instrumentors

# You can combine multiple Traceloop instrumentors
from opentelemetry.instrumentation.openai import OpenAIInstrumentor
 from opentelemetry.instrumentation.anthropic import AnthropicInstrumentor

 # Step 1: Initialize HoneyHive tracer first (without instrumentors)
 tracer = HoneyHiveTracer.init(
     project="your-project"  # Or set HH_PROJECT environment variable
 )

 # Step 2: Initialize instrumentors separately with tracer_provider
 # REPLACE_WITH_INSTRUMENTOR_SETUP
         OpenAIInstrumentor(),      # Works for both OpenAI and Azure OpenAI
         AnthropicInstrumentor()    # Traceloop Anthropic
     ]
 )

Performance Optimization

# Traceloop instrumentors handle batching automatically
# No additional configuration needed for performance

Environment Configuration

# HoneyHive configuration
export HH_API_KEY="your-honeyhive-api-key"
export HH_SOURCE="production"

# Azure OpenAI configuration
export AZURE_OPENAI_API_KEY="your-azure-openai-key"
export AZURE_OPENAI_ENDPOINT="https://your-resource.openai.azure.com/"
export AZURE_OPENAI_API_VERSION="2024-02-01"

# Optional: Traceloop cloud features
export TRACELOOP_API_KEY="your-traceloop-key"
export TRACELOOP_BASE_URL="https://api.traceloop.com"

Comparison: OpenInference vs Traceloop for Azure OpenAI

Feature Comparison
Feature	OpenInference	Traceloop
Setup Complexity	Simple, single instrumentor	Single instrumentor setup
Token Tracking	Basic span attributes	Detailed token metrics + costs
Model Metrics	Model name, basic timing	Cost per model, latency analysis
Performance	Lightweight, fast	Optimized with smart batching
Cost Analysis	Manual calculation needed	Automatic cost per request
Production Ready	✅ Yes	✅ Yes, with cost insights
Debugging	Standard OpenTelemetry	Enhanced LLM-specific debug
Best For	Simple integrations, dev	Production, cost optimization

Migration Between Instrumentors

From OpenInference to Traceloop:

# Before (OpenInference)
from openinference.instrumentation.openai import OpenAIInstrumentor
# Step 1: Initialize HoneyHive tracer first (without instrumentors)
tracer = HoneyHiveTracer.init(
    project="your-project"  # Or set HH_PROJECT environment variable
)

# Step 2: Initialize instrumentor separately with tracer_provider
instrumentor = OpenAIInstrumentor()
instrumentor.instrument(tracer_provider=tracer.provider)

# After (Traceloop) - different instrumentor package
from opentelemetry.instrumentation.openai import OpenAIInstrumentor
# Step 1: Initialize HoneyHive tracer first (without instrumentors)
tracer = HoneyHiveTracer.init(
    project="your-project"  # Or set HH_PROJECT environment variable
)

# Step 2: Initialize instrumentor separately with tracer_provider
instrumentor = OpenAIInstrumentor()
instrumentor.instrument(tracer_provider=tracer.provider)

From Traceloop to OpenInference:

# Before (Traceloop)
from opentelemetry.instrumentation.openai import OpenAIInstrumentor
# Step 1: Initialize HoneyHive tracer first (without instrumentors)
tracer = HoneyHiveTracer.init(
    project="your-project"  # Or set HH_PROJECT environment variable
)

# Step 2: Initialize instrumentor separately with tracer_provider
instrumentor = OpenAIInstrumentor()
instrumentor.instrument(tracer_provider=tracer.provider)

# After (OpenInference)
from openinference.instrumentation.openai import OpenAIInstrumentor
# Step 1: Initialize HoneyHive tracer first (without instrumentors)
tracer = HoneyHiveTracer.init(
    project="your-project"  # Or set HH_PROJECT environment variable
)

# Step 2: Initialize instrumentor separately with tracer_provider
instrumentor = OpenAIInstrumentor()
instrumentor.instrument(tracer_provider=tracer.provider)