Observability in generative AI

In today's AI-driven world, Generative AI Operations (GenAIOps) is revolutionizing how organizations build and deploy intelligent systems. As companies increasingly use AI agents and applications to transform decision-making, enhance customer experiences, and fuel innovation, one element stands paramount: robust evaluation frameworks. Evaluation isn't just a checkpoint. It's the foundation of quality and trust in AI applications. Without rigorous assessment and monitoring, AI systems can produce content that's:

• Fabricated or ungrounded in reality
• Irrelevant or incoherent
• Harmful in perpetuating content risks and stereotypes
• Dangerous in spreading misinformation
• Vulnerable to security exploits

This is where observability becomes essential. These capabilities measure both the frequency and severity of risks in AI outputs, enabling teams to systematically address quality, safety, and security concerns throughout the entire AI development journey—from selecting the right model to monitoring production performance, quality, and safety.

What is observability?

AI observability refers to the ability to monitor, understand, and troubleshoot AI systems throughout their lifecycle. It involves collecting and analyzing signals such as evaluation metrics, logs, traces, and model and agent outputs to gain visibility into performance, quality, safety, and operational health.

What are evaluators?

Evaluators are specialized tools that measure the quality, safety, and reliability of AI responses. By implementing systematic evaluations throughout the AI development lifecycle, teams can identify and address potential issues before they impact users. The following supported evaluators provide comprehensive assessment capabilities across different AI application types and concerns:

General purpose

To learn more, see General purpose evaluators.

Textual similarity

To learn more, see Textual similarity evaluators

RAG (retrieval augmented generation)

To learn more, see Retrieval-augmented Generation (RAG) evaluators.

Safety and security (preview)

To learn more, see Risk and safety evaluators.

Agents (preview)

To learn more, see Agent evaluators.

Azure OpenAI graders (preview)

To learn more, see Azure OpenAI Graders.

Evaluators in the development lifecycle

By using these evaluators strategically throughout the development lifecycle, teams can build more reliable, safe, and effective AI applications that meet user needs while minimizing potential risks.

The three stages of GenAIOps evaluation

GenAIOps uses the following three stages.

Base model selection

Before building your application, you need to select the right foundation. This initial evaluation helps you compare different models based on:

• Quality and accuracy: How relevant and coherent are the model's responses?
• Task performance: Does the model handle your specific use cases efficiently?
• Ethical considerations: Is the model free from harmful biases?
• Safety profile: What is the risk of generating unsafe content?

Tools available: Microsoft Foundry benchmark for comparing models on public datasets or your own data, and the Azure AI Evaluation SDK for testing specific model endpoints.

Preproduction evaluation

After you select a base model, the next step is to develop an AI agent or application. Before you deploy to a production environment, thorough testing is essential to ensure that the AI agent or application is ready for real-world use.

Preproduction evaluation involves:

• Testing with evaluation datasets: These datasets simulate realistic user interactions to ensure the AI agent performs as expected.
• Identifying edge cases: Finding scenarios where the AI agent's response quality might degrade or produce undesirable outputs.
• Assessing robustness: Ensuring that the AI agent can handle a range of input variations without significant drops in quality or safety.
• Measuring key metrics: Metrics such as task adherence, response groundedness, relevance, and safety are evaluated to confirm readiness for production.

The preproduction stage acts as a final quality check, reducing the risk of deploying an AI agent or application that doesn't meet the desired performance or safety standards.

Evaluation Tools and Approaches:

After you get satisfactory results, you can deploy the AI application to production.

Post-production monitoring

• Operational metrics: Regular measurement of key AI agent operational metrics.
• Continuous evaluation: Enables quality and safety evaluation of production traffic at a sampled rate.
• Scheduled evaluation: Enables scheduled quality and safety evaluation using a test dataset to detect drift in the underlying systems.
• Scheduled red teaming: Provides scheduled adversarial testing capabilities to probe for safety and security vulnerabilities.
• Azure Monitor alerts: Swift action when harmful or inappropriate outputs occur. Set up alerts for continuous evaluation to be notified when evaluation results drop below the pass rate threshold in production.

Effective monitoring helps maintain user trust and allows for rapid issue resolution.

Observability provides comprehensive monitoring capabilities essential for today's complex and rapidly evolving AI landscape. Seamlessly integrated with Azure Monitor Application Insights, this solution enables continuous monitoring of deployed AI applications to ensure optimal performance, safety, and quality in production environments.

The Foundry Observability dashboard delivers real-time insights into critical metrics. It allows teams to quickly identify and address performance issues, safety concerns, or quality degradation.

For Agent-based applications, Foundry offers enhanced continuous evaluation capabilities. These capabilities can provide deeper visibility into quality and safety metrics. They can create a robust monitoring ecosystem that adapts to the dynamic nature of AI applications while maintaining high standards of performance and reliability.

By continuously monitoring the AI application's behavior in production, you can maintain high-quality user experiences and swiftly address any issues that surface.

Building trust through systematic evaluation

GenAIOps establishes a reliable process for managing AI applications throughout their lifecycle. By implementing thorough evaluation at each stage—from model selection through deployment and beyond—teams can create AI solutions that aren't just powerful but trustworthy and safe.

Evaluation cheat sheet

Bring your own virtual network for evaluation

For network isolation purposes you can bring your own virtual network for evaluation. To learn more, see How to configure a private link.

Virtual network region support

Region support

Currently certain AI-assisted evaluators are available only in the following regions:

Agent playground evaluation region support

Pricing

Observability features such as Risk and Safety Evaluations and Continuous Evaluations are billed based on consumption as listed in our Azure pricing page.

Related content