Azure AI Foundry

Image provided by Microsoft Learn documentation https://learn.microsoft.com/en-us/azure/ai-studio/concepts/architecture

Introduction

This week at Microsoft Ignite, Azure AI Foundry was unveiled as the rebranded successor to “Azure AI Studio.” This marks a significant step toward unifying AI development tools under one cohesive platform. Azure AI Foundry provides a streamlined toolchain and an SDK designed for efficient consumption of AI models, supporting both OpenAI and Mistral models programmatically. Now accessible at ai.azure.com, Azure AI Foundry retains its core functionality as a collaborative hub, enabling organizations to group resources and manage projects seamlessly. The tool chain is the various services provisioned with Cognitive Services, Prompt Flow and Content Filtering amongst others.

User Experience First Look

Given most developers are switching between IDE and Portals the new look tries to simplify this with everything in one page.

The same navigation pane has expanded and added Safety + Security to reside in one tab. This has also separated the management center that takes a step back for you to quickly identify what is deployed and where.

This separates two panels for the hub and projects so you can get more granularity overall and one place to manage quota which will give you a idea of what is used by region.

Separate areas that have been streamlined in terms of navigation are the Assess and Improve which contain Tracing (preview), Evaluation, Safety + Security.

Tracing on your application which leverages the use of Azure Application Insights and if your using prompt flow this is tied into the SDK if I recall correctly with @trace above your tool call.

The Safety and Security section is cleaned up with your Content Filters (enforced on your deployed models) and Blocklists which you can input what is allowed and not allowed from a language perspective. Its still important to note this is one side of the coin in providing a layer of security as you’ll likely using Azure AI Content Safety as a buffer on top of these features.

If you prefer a playground experience to test out a grouping of prompts you believe can bypass the Prompt Shield you also have the Try it out area that allows a front-end test interface.

Out of the Try it out section the three items still in preview as of this writing are Groundedness detection, Protected Material Detection for Code and Moderate Multimodal Content.

The Model Catalog list as of this writing the available models being 1812 as of today notable that you can also compare with benchmarks to help you with selection this is a differentiator when I think of a platform providing Model-as-service to the end consumer.

For instance if I’m concerned with quality vs cost I can use this comparison chart on various models I’m considering.

Last but not least the Playground has a streamlined interface to help you decide on testing in a controlled UI.

While the Agents was a main theme I haven’t seen any updates on the UI/UX to reflect this topic or API as of yet this will likely change in the coming months. I’d imagine as Agent Frameworks are now taking a first seat in development we will likely see that also being a part of the development workflow such as Autogen, CrewAI, LangGraph.

Future Improvements

It appears the decision to unify more of Copilot Studio with Azure AI Foundry is the bridge given the expansion of how models can be accessed as serverless deployment or hosted on a compute resource for inferencing. Once the Agents portion of the Azure AI Foundry SDK goes General Availability we should be able to use agents programmatically native to the platform which will be a step forward. Region support has broadly expanded from when this service first started so this appears to show the capacity is now where most organization are demanding along with using Azure OpenAI doesn’t have a process workflow for provisioning I was able to deploy Dall-e 3 Model fairly quickly.

Prompt: An astronaut, donned in his space gear, is seen riding a camel. The setting is a vast desert, exuding shades of golden sand as an abstract painting

Given the access to proprietary models such as OpenAI and the expansion of the catalog to have other AI models hosted such as Bria 2, NTT Data this will likely capture a lead in areas similar AWS Bedrock and Vertex AI Studio in the on-going platform wars of building AI applications.

Final Thoughts

Given most of the changes are unifying various portals in one this is a step to bridge the legacy Open AI Studio endpoint and AI Studio to one which help stop any confusion. The messaging behind unifying Cognitive Services when you deploy to Azure AI Foundry helps with various resources under one area and the management center helps with identifying quotas and regional constraints if they do exist. Azure Policy is also leveraging custom policies for governance in the realm of ‘Allowed Models’ for deployment with a template but you can also add and modify as needed. This product will likely take a front seat in the next twelve months and its great to see the continued investment in simplifying the building of applications with more templates out of the box hosted on Github.

For those who have purchased my course via Udemy on Azure AI Studio the changes of this Azure AI Foundry will start to roll out relatively soon as I’m placing the content and updates accordingly.

Resources: https://learn.microsoft.com/en-us/azure/ai-studio/what-is-ai-studio