More advanced scenarios for telemetry
Note
This article will use Aspire Dashboard for illustration. If you prefer to use other tools, please refer to the documentation of the tool you are using on setup instructions.
Auto Function Calling
Auto Function Calling is a Semantic Kernel feature that allows the kernel to automatically execute functions when the model responds with function calls, and provide the results back to the model. This feature is useful for scenarios where a query requires multiple iterations of function calls to get a final natural language response. For more details, please see these GitHub samples.
Note
Function calling is not supported by all models.
Tip
You will hear the term "tools" and "tool calling" sometimes used interchangeably with "functions" and "function calling".
Prerequisites
- An Azure OpenAI chat completion deployment that supports function calling.
- Docker
- The latest .Net SDK for your operating system.
- An Azure OpenAI chat completion deployment that supports function calling.
- Docker
- Python 3.10, 3.11, or 3.12 installed on your machine.
Note
Semantic Kernel Observability is not yet available for Java.
Setup
Create a new console application
In a terminal, run the following command to create a new console application in C#:
dotnet new console -n TelemetryAutoFunctionCallingQuickstart
Navigate to the newly created project directory after the command completes.
Install required packages
Semantic Kernel
dotnet add package Microsoft.SemanticKernelOpenTelemetry Console Exporter
dotnet add package OpenTelemetry.Exporter.OpenTelemetryProtocol
Create a simple application with Semantic Kernel
From the project directory, open the Program.cs file with your favorite editor. We are going to create a simple application that uses Semantic Kernel to send a prompt to a chat completion model. Replace the existing content with the following code and fill in the required values for deploymentName, endpoint, and apiKey:
using System.ComponentModel;
using Microsoft.Extensions.DependencyInjection;
using Microsoft.Extensions.Logging;
using Microsoft.SemanticKernel;
using Microsoft.SemanticKernel.Connectors.OpenAI;
using OpenTelemetry;
using OpenTelemetry.Logs;
using OpenTelemetry.Metrics;
using OpenTelemetry.Resources;
using OpenTelemetry.Trace;
namespace TelemetryAutoFunctionCallingQuickstart
{
class BookingPlugin
{
[KernelFunction("FindAvailableRooms")]
[Description("Finds available conference rooms for today.")]
public async Task<List<string>> FindAvailableRoomsAsync()
{
// Simulate a remote call to a booking system.
await Task.Delay(1000);
return ["Room 101", "Room 201", "Room 301"];
}
[KernelFunction("BookRoom")]
[Description("Books a conference room.")]
public async Task<string> BookRoomAsync(string room)
{
// Simulate a remote call to a booking system.
await Task.Delay(1000);
return $"Room {room} booked.";
}
}
class Program
{
static async Task Main(string[] args)
{
// Endpoint to the Aspire Dashboard
var endpoint = "http://localhost:4317";
var resourceBuilder = ResourceBuilder
.CreateDefault()
.AddService("TelemetryAspireDashboardQuickstart");
// Enable model diagnostics with sensitive data.
AppContext.SetSwitch("Microsoft.SemanticKernel.Experimental.GenAI.EnableOTelDiagnosticsSensitive", true);
using var traceProvider = Sdk.CreateTracerProviderBuilder()
.SetResourceBuilder(resourceBuilder)
.AddSource("Microsoft.SemanticKernel*")
.AddOtlpExporter(options => options.Endpoint = new Uri(endpoint))
.Build();
using var meterProvider = Sdk.CreateMeterProviderBuilder()
.SetResourceBuilder(resourceBuilder)
.AddMeter("Microsoft.SemanticKernel*")
.AddOtlpExporter(options => options.Endpoint = new Uri(endpoint))
.Build();
using var loggerFactory = LoggerFactory.Create(builder =>
{
// Add OpenTelemetry as a logging provider
builder.AddOpenTelemetry(options =>
{
options.SetResourceBuilder(resourceBuilder);
options.AddOtlpExporter(options => options.Endpoint = new Uri(endpoint));
// Format log messages. This is default to false.
options.IncludeFormattedMessage = true;
options.IncludeScopes = true;
});
builder.SetMinimumLevel(LogLevel.Information);
});
IKernelBuilder builder = Kernel.CreateBuilder();
builder.Services.AddSingleton(loggerFactory);
builder.AddAzureOpenAIChatCompletion(
deploymentName: "your-deployment-name",
endpoint: "your-azure-openai-endpoint",
apiKey: "your-azure-openai-api-key"
);
builder.Plugins.AddFromType<BookingPlugin>();
Kernel kernel = builder.Build();
var answer = await kernel.InvokePromptAsync(
"Reserve a conference room for me today.",
new KernelArguments(
new OpenAIPromptExecutionSettings {
ToolCallBehavior = ToolCallBehavior.AutoInvokeKernelFunctions
}
)
);
Console.WriteLine(answer);
}
}
}
In the code above, we first define a mock conference room booking plugin with two functions: FindAvailableRoomsAsync and BookRoomAsync. We then create a simple console application that registers the plugin to the kernel, and ask the kernel to automatically call the functions when needed.
Create a new Python virtual environment
python -m venv telemetry-auto-function-calling-quickstart
Activate the virtual environment.
telemetry-auto-function-calling-quickstart\Scripts\activate
Install required packages
pip install semantic-kernel opentelemetry-exporter-otlp-proto-grpc
Create a simple Python script with Semantic Kernel
Create a new Python script and open it with your favorite editor.
New-Item -Path telemetry_auto_function_calling_quickstart.py -ItemType file
We are going to create a simple Python script that uses Semantic Kernel to send a prompt to a chat completion model. Replace the existing content with the following code and fill in the required values for deployment_name, endpoint, and api_key:
import asyncio
import logging
from typing import Annotated
from opentelemetry._logs import set_logger_provider
from opentelemetry.exporter.otlp.proto.grpc._log_exporter import OTLPLogExporter
from opentelemetry.exporter.otlp.proto.grpc.metric_exporter import OTLPMetricExporter
from opentelemetry.exporter.otlp.proto.grpc.trace_exporter import OTLPSpanExporter
from opentelemetry.metrics import set_meter_provider
from opentelemetry.sdk._logs import LoggerProvider, LoggingHandler
from opentelemetry.sdk._logs.export import BatchLogRecordProcessor
from opentelemetry.sdk.metrics import MeterProvider
from opentelemetry.sdk.metrics.export import PeriodicExportingMetricReader
from opentelemetry.sdk.metrics.view import DropAggregation, View
from opentelemetry.sdk.resources import Resource
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.export import BatchSpanProcessor
from opentelemetry.semconv.resource import ResourceAttributes
from opentelemetry.trace import set_tracer_provider
from semantic_kernel import Kernel
from semantic_kernel.connectors.ai.function_choice_behavior import FunctionChoiceBehavior
from semantic_kernel.connectors.ai.open_ai import AzureChatCompletion
from semantic_kernel.connectors.ai.prompt_execution_settings import PromptExecutionSettings
from semantic_kernel.functions.kernel_arguments import KernelArguments
from semantic_kernel.functions.kernel_function_decorator import kernel_function
class BookingPlugin:
@kernel_function(
name="find_available_rooms",
description="Find available conference rooms for today.",
)
def find_available_rooms(self,) -> Annotated[list[str], "A list of available rooms."]:
return ["Room 101", "Room 201", "Room 301"]
@kernel_function(
name="book_room",
description="Book a conference room.",
)
def book_room(self, room: str) -> Annotated[str, "A confirmation message."]:
return f"Room {room} booked."
# Endpoint to the Aspire Dashboard
endpoint = "http://localhost:4317"
# Create a resource to represent the service/sample
resource = Resource.create({ResourceAttributes.SERVICE_NAME: "telemetry-aspire-dashboard-quickstart"})
def set_up_logging():
exporter = OTLPLogExporter(endpoint=endpoint)
# Create and set a global logger provider for the application.
logger_provider = LoggerProvider(resource=resource)
# Log processors are initialized with an exporter which is responsible
# for sending the telemetry data to a particular backend.
logger_provider.add_log_record_processor(BatchLogRecordProcessor(exporter))
# Sets the global default logger provider
set_logger_provider(logger_provider)
# Create a logging handler to write logging records, in OTLP format, to the exporter.
handler = LoggingHandler()
# Add filters to the handler to only process records from semantic_kernel.
handler.addFilter(logging.Filter("semantic_kernel"))
# Attach the handler to the root logger. `getLogger()` with no arguments returns the root logger.
# Events from all child loggers will be processed by this handler.
logger = logging.getLogger()
logger.addHandler(handler)
logger.setLevel(logging.INFO)
def set_up_tracing():
exporter = OTLPSpanExporter(endpoint=endpoint)
# Initialize a trace provider for the application. This is a factory for creating tracers.
tracer_provider = TracerProvider(resource=resource)
# Span processors are initialized with an exporter which is responsible
# for sending the telemetry data to a particular backend.
tracer_provider.add_span_processor(BatchSpanProcessor(exporter))
# Sets the global default tracer provider
set_tracer_provider(tracer_provider)
def set_up_metrics():
exporter = OTLPMetricExporter(endpoint=endpoint)
# Initialize a metric provider for the application. This is a factory for creating meters.
meter_provider = MeterProvider(
metric_readers=[PeriodicExportingMetricReader(exporter, export_interval_millis=5000)],
resource=resource,
views=[
# Dropping all instrument names except for those starting with "semantic_kernel"
View(instrument_name="*", aggregation=DropAggregation()),
View(instrument_name="semantic_kernel*"),
],
)
# Sets the global default meter provider
set_meter_provider(meter_provider)
# This must be done before any other telemetry calls
set_up_logging()
set_up_tracing()
set_up_metrics()
async def main():
# Create a kernel and add a service
kernel = Kernel()
kernel.add_service(AzureChatCompletion(
api_key="your-azure-openai-api-key",
endpoint="your-azure-openai-endpoint",
deployment_name="your-deployment-name"
))
kernel.add_plugin(BookingPlugin(), "BookingPlugin")
answer = await kernel.invoke_prompt(
"Reserve a conference room for me today.",
arguments=KernelArguments(
settings=PromptExecutionSettings(
function_choice_behavior=FunctionChoiceBehavior.Auto(),
),
),
)
print(answer)
if __name__ == "__main__":
asyncio.run(main())
In the code above, we first define a mock conference room booking plugin with two functions: find_available_rooms and book_room. We then create a simple Python script that registers the plugin to the kernel, and ask the kernel to automatically call the functions when needed.
Environment variables
Please refer to this article for more information on setting up the required environment variables to enable the kernel to emit spans for AI connectors.
Note
Semantic Kernel Observability is not yet available for Java.
Start the Aspire Dashboard
Follow the instructions here to start the dashboard. Once the dashboard is running, open a browser and navigate to http://localhost:18888 to access the dashboard.
Run
Run the console application with the following command:
dotnet run
Run the Python script with the following command:
python telemetry_auto_function_calling_quickstart.py
Note
Semantic Kernel Observability is not yet available for Java.
You should see an output similar to the following:
Room 101 has been successfully booked for you today.
Inspect telemetry data
After running the application, head over to the dashboard to inspect the telemetry data.
Find the trace for the application in the Traces tab. You should five spans in the trace:
These 5 spans represent the internal operations of the kernel with auto function calling enabled. It first invokes the model, which requests a function call. Then the kernel automatically executes the function FindAvailableRoomsAsync and returns the result to the model. The model then requests another function call to make a reservation, and the kernel automatically executes the function BookRoomAsync and returns the result to the model. Finally, the model returns a natural language response to the user.
And if you click on the last span, and look for the prompt in the gen_ai.content.prompt event, you should see something similar to the following:
[
{ "role": "user", "content": "Reserve a conference room for me today." },
{
"role": "Assistant",
"content": null,
"tool_calls": [
{
"id": "call_NtKi0OgOllJj1StLkOmJU8cP",
"function": { "arguments": {}, "name": "FindAvailableRooms" },
"type": "function"
}
]
},
{
"role": "tool",
"content": "[\u0022Room 101\u0022,\u0022Room 201\u0022,\u0022Room 301\u0022]"
},
{
"role": "Assistant",
"content": null,
"tool_calls": [
{
"id": "call_mjQfnZXLbqp4Wb3F2xySds7q",
"function": { "arguments": { "room": "Room 101" }, "name": "BookRoom" },
"type": "function"
}
]
},
{ "role": "tool", "content": "Room Room 101 booked." }
]
This is the chat history that gets built up as the model and the kernel interact with each other. This is sent to the model in the last iteration to get a natural language response.
Find the trace for the application in the Traces tab. You should five spans in the trace grouped under the AutoFunctionInvocationLoop span:
These 5 spans represent the internal operations of the kernel with auto function calling enabled. It first invokes the model, which requests a function call. Then the kernel automatically executes the function find_available_rooms and returns the result to the model. The model then requests another function call to make a reservation, and the kernel automatically executes the function book_room and returns the result to the model. Finally, the model returns a natural language response to the user.
And if you click on the last span, and look for the prompt in the gen_ai.content.prompt event, you should see something similar to the following:
[
{ "role": "user", "content": "Reserve a conference room for me today." },
{
"role": "assistant",
"tool_calls": [
{
"id": "call_ypqO5v6uTRlYH9sPTjvkGec8",
"type": "function",
"function": {
"name": "BookingPlugin-find_available_rooms",
"arguments": "{}"
}
}
]
},
{
"role": "tool",
"content": "['Room 101', 'Room 201', 'Room 301']",
"tool_call_id": "call_ypqO5v6uTRlYH9sPTjvkGec8"
},
{
"role": "assistant",
"tool_calls": [
{
"id": "call_XDZGeTfNiWRpYKoHoH9TZRoX",
"type": "function",
"function": {
"name": "BookingPlugin-book_room",
"arguments": "{\"room\":\"Room 101\"}"
}
}
]
},
{
"role": "tool",
"content": "Room Room 101 booked.",
"tool_call_id": "call_XDZGeTfNiWRpYKoHoH9TZRoX"
}
]
This is the chat history that gets built up as the model and the kernel interact with each other. This is sent to the model in the last iteration to get a natural language response.
Note
Semantic Kernel Observability is not yet available for Java.
Error handling
If an error occurs during the execution of a function, the kernel will automatically catch the error and return an error message to the model. The model can then use this error message to provide a natural language response to the user.
Modify the BookRoomAsync function in the C# code to simulate an error:
[KernelFunction("BookRoom")]
[Description("Books a conference room.")]
public async Task<string> BookRoomAsync(string room)
{
// Simulate a remote call to a booking system.
await Task.Delay(1000);
throw new Exception("Room is not available.");
}
Run the application again and observe the trace in the dashboard. You should see the span representing the kernel function call with an error:
Note
It is very likely that the model responses to the error may vary each time you run the application, because the model is stochastic. You may see the model reserving all three rooms at the same time, or reserving one the first time then reserving the other two the second time, etc.
Modify the book_room function in the Python code to simulate an error:
@kernel_function(
name="book_room",
description="Book a conference room.",
)
async def book_room(self, room: str) -> Annotated[str, "A confirmation message."]:
# Simulate a remote call to a booking system
await asyncio.sleep(1)
raise Exception("Room is not available.")
Run the application again and observe the trace in the dashboard. You should see the span representing the kernel function call with an error and the stack trace:
Note
It is very likely that the model responses to the error may vary each time you run the application, because the model is stochastic. You may see the model reserving all three rooms at the same time, or reserving one the first time then reserving the other two the second time, etc.
Note
Semantic Kernel Observability is not yet available for Java.
Next steps and further reading
In production, your services may get a large number of requests. Semantic Kernel will generate a large amount of telemetry data. some of which may not be useful for your use case and will introduce unnecessary costs to store the data. You can use the sampling feature to reduce the amount of telemetry data that is collected.
Observability in Semantic Kernel is constantly improving. You can find the latest updates and new features in the GitHub repository.