Structured outputs in OpenAIAssistantAgent?

[StefandewitSPS]

When will structured outputs be available in OpenAIAssistantAgent? :)

[moonbox3]

Hi @StefandewitSPS, thanks for your question. Are you using SK Python or .Net? For Python, it will be available next week -- we're going to make some updates to the agent, similar to what .Net just did, to simplify its construction and allow devs more flexibility with how the assistant agent is created. I believe the .Net changes just went into main today, so they'll be included during the next release (next week as well).

[StefandewitSPS]

Thanks for the update regarding structured outputs, I thought. While I understood the general concept, I still had one pressing question: What’s the best way to implement this in my application?

My system is designed around multiple agents, each with its own set of plugins. These plugins handle numerous functions, and my ultimate goal is to return structured layouts to the client. For instance, I want to provide:

GridLayout – a grid containing multiple results
ButtonLayout – a button triggering an action
… and potentially more dynamic layouts in the future.
But the real challenge lies in how to structure these outputs efficiently. Should my kernel functions return JSON, explicitly defining the layout type within the response? Or is it better for the kernel functions to return a class, which can then be serialized accordingly?

Ideally, I want a consistent JSON format that looks like this:

JSON:
{
"Layout": "Enum(GridLayout, ButtonLayout, etc.)",
"Data": [],
"Class": "The type of the class"
}
Since I’m using the .NET version of Semantic Kernel, I’m eager to find the best approach. Is there a recommended example or best practice to follow? Any guidance would be greatly appreciated as I aim to build a scalable and maintainable solution.

[moonbox3]

Tagging @dmytrostruk for .Net structured outputs help.

[dmytrostruk]

But the real challenge lies in how to structure these outputs efficiently. Should my kernel functions return JSON, explicitly defining the layout type within the response? Or is it better for the kernel functions to return a class, which can then be serialized accordingly?

@StefandewitSPS I think your kernel function can return a class that will describe a layout. With class you will have easy access to specific property, so it will be easier for you to test how your function works, comparing to returning a JSON string. It will be also easier to validate your layout result if needed before returning it back to the client. General recommendation - use strong types until you get to the point when you need to pass your information to the other component in your solution, such as AI model or front-end client.