Access context from an AIFunction

[verdie-g]

We are migrating from PydanticAI to MEAI and we are facing an issue when trying to get some context from a tool.

With PydanticAI you can do this

class MyContext(BaseModel):
    user_id: int
    country: str

# Tool that uses context
@Tool
def get_user_country(ctx: RunContext[MyContext]) -> str:
    return f"User {ctx.user_id} is from {ctx.country}"

agent = Agent(
    "gpt-4.1-mini",
    tools=[get_user_country],
)

# Inject context
result = agent.run(
    "Where is the user from?",
    context=MyContext(user_id=42, country="France")
)

but I couldn't find an equivalent with MEAI.

The work-around was to create a UserContextAccessor à la IHttpContextAccessor.

Another work-around is to create the functions on the fly to capture the context in the closure but it's very inconvenient.

Is there a better way or would you consider adding a new API to do that?

[stephentoub]

I'm unclear what you're asking for specifically. There are many ways you can get data into an AIFunction invocation.

You can associate data with the function at the time it's created, either via closure or via AdditionalProperties, which if you're creating your own you can populate however you want, or using AIFunctionFactory you can populate using AIFunctionFactoryOptions.AdditionalProperties.

When invoking an AIFunction, you pass in AIFunctionArguments, which not only includes the dictionary of named arguments, it also includes an arbitrary IDictionary<object, object> for passing additional ambient data into the function. That AIFunctionArguments also includes an IServiceProvider, so that the AIFunction can resolve arbitrary state from a dependency injection container. If you want to access either the AIFunctionArguments or the IServiceProvider from within your AIFunctionFactory.Create delegate, just create a parameter typed as IServiceProvider or AIFunctionArguments, and that parameter will be bound to the relevant data from the AIFunctionArguments passed to AIFunction.InvokeAsync.

If you want more fine-grained control over how parameters are bound, you can utilize the AIFunctionFactoryOptions.ConfigureParameterBinding, which puts you in control over how each parameter is populated. The MCP C# SDK, for example, uses that to automatically bind parameters from DI:
https://github.com/modelcontextprotocol/csharp-sdk/blob/d344c651203841ec1c9e828736d234a6e4aebd07/src/ModelContextProtocol.Core/Server/AIFunctionMcpServerTool.cs#L83-L107

If you use the AIFunctionFactory.Create overload that takes an instance MethodInfo and a createInstanceFunc, you can also run your own arbitrary logic when creating the target object the instance method will be called on, each time, and you can do whatever you want to configure that instance.

Is there something more you're hoping for?

[verdie-g]

Thanks for the detailed answer. Let me give more context

You can associate data with the function at the time it's created

I want to access data that is scoped to the request.

When invoking an AIFunction, you pass in AIFunctionArguments

I'm never invoking an AIFunction manually, I let the FunctionInvokingChatClient take care of that. Is there a use-case for calling AIFunction.InvokeAsync? 🤔

If you want more fine-grained control over how parameters are bound, you can utilize the AIFunctionFactoryOptions.ConfigureParameterBinding

I'll have a look but this sounds more advanced.

If you use the AIFunctionFactory.Create overload that takes an instance MethodInfo and a createInstanceFunc

Doesn't look super convenient either.

Is there something more you're hoping for?

Ideally, I would do

private readonly AIFunction _getCurrentUser =
    AIFunctionFactory.Create((MyContext ctx) => ctx.UserId, "get_current_user");

var response = await chatClient.GetResponseAsync(
    "Who is the current user?",
    new ChatOptions { Tools = [_getCurrentUser] },
    new MyContext { UserId = currentRequest.UserId });

a Dictionary of objects would work too.

[stephentoub]

I'm never invoking an AIFunction manually, I let the FunctionInvokingChatClient take care of that

FICC configures an AIFunctionArguments that it passes into the AIFunction. As noted, that AIFunctionArguments includes not only the arguments to be provided to the function, but also the IServiceProvider that the FICC was itself provided with. So if you constructed your FICC using standard DI means, that IServiceProvider will be passed all the way into your AIFunction, at which point you can query it for anything you want from DI. So, if for example "currentRequest" is itself actually coming from DI, the AIFunction itself can just take an IServiceProvider argument, and do services.GetRequiredService to resolve the data from the DI container.

a Dictionary of objects would work too.

FICC publishes a bunch of state about the current invocation to FunctionInvokingChatClient.CurrentContext, including not only the arguments, but all of the input ChatMessages, the ChatOptions, details on which function is being invoked (out of how many), etc. If as you've stated you're only concerned with when your function is invoked by FICC, then you can add whatever you want into ChatOptions.AdditionalProperties, and extract that inside of your AIFunction from FunctionInvokingChatClient.CurrentContext.Options.AdditionalProperties.

[verdie-g]

var getCurrentUser = AIFunctionFactory.Create(
    () => FunctionInvokingChatClient.CurrentContext!.Options!.AdditionalProperties!["user.id"],
    "get_current_user");
var response = await chatClient.GetResponseAsync(
    "Who is the current user?",
    new ChatOptions
    {
        Tools = [getCurrentUser],
        AdditionalProperties = new AdditionalPropertiesDictionary { ["user.id"] = 123 },
    });

is working perfectly thank you!

I've open dotnet/docs#48469 to update the doc.