Does the C# SDK generate an output schema for return types when using AIFunctionFactory.Create()?

[major-mayer]

I'm using the Microsoft Agent Framework (C# SDK) and creating function tools via AIFunctionFactory.Create.
The documentation clearly explains how input parameter schemas are generated, but it is not clear whether the return type is used to generate an output schema that is exposed to the LLM.
I only specifies that "Return values are serialized to JsonElement using serializerOptions if providedReturn values are serialized to JsonElement using serializerOptions if provided...".

Right now, we are returning all our function return values manually converted into JSON strings, and I wonder if it's worth the effort to switch to returning the actual types that our backend returns.
My hope is that we can then use the Description attributes to explain the model what kind of return it can expect before (and after) executing the function tool, similar to how things work with input parameters.

Something like:

[Description("Description of the result object in general")]
   public class TransactionResult
   {
       [Description("Description of individual properties")]
       public string TransactionCode { get; set; }
   }

So my questions are specifically:

1. Does the framework generate a JSON schema for the return type and does this include the Description attributes of the class fields?
2. Is this schema included in the tool definition sent to the model?
3. If yes, why is this not documented or am I missing something?
4. If no, what is the recommended way to communicate output structure for function tools in C#?

[stephentoub]

Does the framework generate a JSON schema for the return type

Yes, e.g. this:

using System.ComponentModel;
using Microsoft.Extensions.AI;

AIFunction f = AIFunctionFactory.Create(() => new TransactionResult());
Console.WriteLine(f.ReturnJsonSchema);

[Description("Description of the result object in general")] 
class TransactionResult
{
    [Description("Description of individual properties")]
    public string TransactionCode { get; set; }
}

outputs this:

{"description":"Description of the result object in general","type":"object","properties":{"transactionCode":{"description":"Description of individual properties","type":"string"}}}

Is this schema included in the tool definition sent to the model?

That's entirely up to the client implementation and protocol used by whatever AI service is being utilized. Most of them don't actually support any structured way for sending function return schema, only function parameter schema, e.g. here's what OpenAI accepts:

The best you could do with OpenAI would be to just put the schema into the description and hope that it somehow factored that in.

If yes, why is this not documented or am I missing something?

Why is what not documented?

[major-mayer]

Hey @stephentoub
Thanks for your response.
Good to know that the framework generates the JSON Schema for the returned values as well.
Seems like i've missed the ReturnJsonSchema field.

I will try to automatically include the return schema in the description then.

Why is what not documented?

In my opinion this could be noted for example at this page: https://learn.microsoft.com/en-us/agent-framework/agents/tools/function-tools?pivots=programming-language-csharp
It only talks about input parameters but not about the output (schema) that tools produces and how to provide this to the model.