Overview

Monitor, troubleshoot, and evaluate your LLM-powered applications, such as chatbots or data extraction tools, using Azure OpenAI.

If you are building LLM applications, use LLM Observability to investigate the root cause of issues, monitor operational performance, and evaluate the quality, privacy, and safety of your LLM applications.

See the LLM Observability tracing view video for an example of how you can investigate a trace.

Azure OpenAI enables development of copilots and generative AI applications using OpenAI’s library of models. Use the Datadog integration to track the performance and usage of the Azure OpenAI API and deployments.

Setup

LLM Observability: End-to-end visibility into your LLM application using Azure OpenAI

You can enable LLM Observability in different environments. Follow the appropriate setup based on your scenario:

Installation for Python

If you do not have the Datadog Agent:

1. Install the ddtrace package:

  pip install ddtrace

2. Start your application with the following command, enabling Agentless mode:

  DD_SITE=<YOUR_DATADOG_SITE> DD_API_KEY=<YOUR_API_KEY> DD_LLMOBS_ENABLED=1 DD_LLMOBS_AGENTLESS_ENABLED=1 DD_LLMOBS_ML_APP=<YOUR_ML_APP_NAME> ddtrace-run python <YOUR_APP>.py

If you already have the Datadog Agent installed:

1. Make sure the Agent is running and that APM and StatsD are enabled. For example, use the following command with Docker:

docker run -d \
  --cgroupns host \
  --pid host \
  -v /var/run/docker.sock:/var/run/docker.sock:ro \
  -v /proc/:/host/proc/:ro \
  -v /sys/fs/cgroup/:/host/sys/fs/cgroup:ro \
  -e DD_API_KEY=<DATADOG_API_KEY> \
  -p 127.0.0.1:8126:8126/tcp \
  -p 127.0.0.1:8125:8125/udp \
  -e DD_DOGSTATSD_NON_LOCAL_TRAFFIC=true \
  -e DD_APM_ENABLED=true \
  gcr.io/datadoghq/agent:latest

2. If you haven’t already, install the ddtrace package:

  pip install ddtrace

3. Start your application using the ddtrace-run command to automatically enable tracing:

   DD_SITE=<YOUR_DATADOG_SITE> DD_API_KEY=<YOUR_API_KEY> DD_LLMOBS_ENABLED=1 DD_LLMOBS_ML_APP=<YOUR_ML_APP_NAME> ddtrace-run python <YOUR_APP>.py

Note: If the Agent is running on a custom host or port, set DD_AGENT_HOST and DD_TRACE_AGENT_PORT accordingly.

If you are running LLM Observability in a serverless environment (Azure Functions):

Enable LLM Observability by setting the following environment variables:

DD_SITE=<YOUR_DATADOG_SITE> DD_API_KEY=<YOUR_API_KEY> DD_LLMOBS_ENABLED=1 DD_LLMOBS_ML_APP=<YOUR_ML_APP_NAME>

Note: In serverless environments, Datadog automatically flushes spans when the Azure function finishes running.

Automatic Azure OpenAI tracing

The Azure OpenAI integration is automatically enabled when LLM Observability is configured. This captures latency, errors, input and output messages, as well as token usage for Azure OpenAI calls.

The following methods are traced for both synchronous and asynchronous Azure OpenAI operations:

• AzureOpenAI().completions.create()
• AsyncAzureOpenAI().completions.create()
• AzureOpenAI().chat.completions.create()
• AsyncAzureOpenAI().chat.completions.create()

No additional setup is required for these methods.

Validation

Validate that LLM Observability is properly capturing spans by checking your application logs for successful span creation. You can also run the following command to check the status of the ddtrace integration:

ddtrace-run --info

Look for the following message to confirm the setup:

Agent error: None

Debugging

If you encounter issues during setup, enable debug logging by passing the --debug flag:

ddtrace-run --debug

This displays any errors related to data transmission or instrumentation, including issues with Azure OpenAI traces.

Infrastructure Monitoring: Metrics and visibility on your Azure OpenAI resources

If you haven’t already, set up the Microsoft Azure integration first. There are no other installation steps.

Data Collected

Metrics

Service Checks

The Azure OpenAI integration does not include any service checks.

Events

The Azure OpenAI integration does not include any events.

Troubleshooting

Need help? Contact Datadog support.