Python: Add arguments and results attributes to execute tool span

python/samples/getting_started_with_agents/multi_agent_orchestration/step4c_handoff_mix_agent_types.py

@@ -4,6 +4,7 @@
 
 from azure.ai.projects.aio import AIProjectClient
 from azure.identity.aio import DefaultAzureCredential
+from opentelemetry.trace import NoOpTracerProvider
 
 from samples.getting_started_with_agents.multi_agent_orchestration.observability import enable_observability
 from semantic_kernel.agents import (
@@ -197,7 +198,7 @@ async def main():
     )
 
     # 2. Create a runtime and start it
-    runtime = InProcessRuntime()
+    runtime = InProcessRuntime(tracer_provider=NoOpTracerProvider())
     runtime.start()
 
     try:

python/semantic_kernel/functions/kernel_arguments.py

@@ -1,7 +1,10 @@
 # Copyright (c) Microsoft. All rights reserved.
 
+import json
 from typing import TYPE_CHECKING, Any
 
+from pydantic import BaseModel
+
 from semantic_kernel.const import DEFAULT_SERVICE_NAME
 
 if TYPE_CHECKING:
@@ -103,3 +106,17 @@ def __ior__(self, value: "SupportsKeysAndGetItem[Any, Any] | Iterable[tuple[Any,
                 self.execution_settings = value.execution_settings.copy()
 
         return self
+
+    def dumps(self, include_execution_settings: bool = False) -> str:
+        """Serializes the KernelArguments to a JSON string."""
+        data = dict(self)
+        if include_execution_settings and self.execution_settings:
+            data["execution_settings"] = self.execution_settings
+
+        def default(obj):
+            if isinstance(obj, BaseModel):
+                return obj.model_dump()
+
+            return str(obj)
+
+        return json.dumps(data, default=default)

python/semantic_kernel/functions/kernel_function.py

@@ -1,5 +1,6 @@
 # Copyright (c) Microsoft. All rights reserved.
 
+import json
 import logging
 import time
 from abc import abstractmethod
@@ -32,6 +33,7 @@
 from semantic_kernel.prompt_template.kernel_prompt_template import KernelPromptTemplate
 from semantic_kernel.prompt_template.prompt_template_base import PromptTemplateBase
 from semantic_kernel.utils.telemetry.model_diagnostics import function_tracer
+from semantic_kernel.utils.telemetry.model_diagnostics.gen_ai_attributes import TOOL_CALL_ARGUMENTS, TOOL_CALL_RESULT
 
 if TYPE_CHECKING:
     from semantic_kernel.connectors.ai.prompt_execution_settings import PromptExecutionSettings
@@ -260,6 +262,9 @@ async def invoke(
             KernelFunctionLogMessages.log_function_invoking(logger, self.fully_qualified_name)
             KernelFunctionLogMessages.log_function_arguments(logger, arguments)
 
+            if function_tracer.are_sensitive_events_enabled():
+                current_span.set_attribute(TOOL_CALL_ARGUMENTS, arguments.dumps())
+
             attributes = {MEASUREMENT_FUNCTION_TAG_NAME: self.fully_qualified_name}
             starting_time_stamp = time.perf_counter()
             try:
@@ -272,6 +277,13 @@ async def invoke(
                 KernelFunctionLogMessages.log_function_invoked_success(logger, self.fully_qualified_name)
                 KernelFunctionLogMessages.log_function_result_value(logger, function_context.result)
 
+                if function_tracer.are_sensitive_events_enabled():
+                    try:
+                        result = str(function_context.result.value) if function_context.result else None
+                    except Exception as e:
+                        result = str(e)
+                    current_span.set_attribute(TOOL_CALL_RESULT, result)
+
                 return function_context.result
             except Exception as e:
                 self._handle_exception(current_span, e, attributes)
@@ -322,6 +334,9 @@ async def invoke_stream(
             KernelFunctionLogMessages.log_function_streaming_invoking(logger, self.fully_qualified_name)
             KernelFunctionLogMessages.log_function_arguments(logger, arguments)
 
+            if function_tracer.are_sensitive_events_enabled():
+                current_span.set_attribute(TOOL_CALL_ARGUMENTS, arguments.dumps())
+
             attributes = {MEASUREMENT_FUNCTION_TAG_NAME: self.fully_qualified_name}
             starting_time_stamp = time.perf_counter()
             try:
@@ -331,15 +346,27 @@ async def invoke_stream(
                 )
                 await stack(function_context)
 
+                function_results: list[Any] = []
                 if function_context.result is not None:
                     if isasyncgen(function_context.result.value):
                         async for partial in function_context.result.value:
+                            function_results.append(partial)
                             yield partial
                     elif isgenerator(function_context.result.value):
                         for partial in function_context.result.value:
+                            function_results.append(partial)
                             yield partial
                     else:
+                        function_results.append(function_context.result.value)
                         yield function_context.result
+
+                if function_tracer.are_sensitive_events_enabled():
+                    results: list[str] = []
+                    try:
+                        results.append(str(function_results))
+                    except Exception as e:
+                        results.append(str(e))
+                    current_span.set_attribute(TOOL_CALL_RESULT, json.dumps(results))
             except Exception as e:
                 self._handle_exception(current_span, e, attributes)
                 raise e

python/semantic_kernel/utils/telemetry/model_diagnostics/function_tracer.py

@@ -4,12 +4,14 @@
 
 from opentelemetry import trace
 
+from semantic_kernel.utils.feature_stage_decorator import experimental
 from semantic_kernel.utils.telemetry.model_diagnostics.gen_ai_attributes import (
     OPERATION,
     TOOL_CALL_ID,
     TOOL_DESCRIPTION,
     TOOL_NAME,
 )
+from semantic_kernel.utils.telemetry.model_diagnostics.model_diagnostics_settings import ModelDiagnosticSettings
 
 if TYPE_CHECKING:
     from semantic_kernel.functions.kernel_function import KernelFunction
@@ -19,6 +21,20 @@
 # https://opentelemetry.io/docs/specs/semconv/gen-ai/gen-ai-spans/#execute-tool-span
 OPERATION_NAME = "execute_tool"
 
+# To enable these features, set one of the following environment variables to true:
+#    SEMANTICKERNEL_EXPERIMENTAL_GENAI_ENABLE_OTEL_DIAGNOSTICS
+#    SEMANTICKERNEL_EXPERIMENTAL_GENAI_ENABLE_OTEL_DIAGNOSTICS_SENSITIVE
+MODEL_DIAGNOSTICS_SETTINGS = ModelDiagnosticSettings()
+
+
+@experimental
+def are_sensitive_events_enabled() -> bool:
+    """Check if sensitive events are enabled.
+
+    Sensitive events are enabled if the diagnostic with sensitive events is enabled.
+    """
+    return MODEL_DIAGNOSTICS_SETTINGS.enable_otel_diagnostics_sensitive
+
 
 def start_as_current_span(
     tracer: trace.Tracer,

python/semantic_kernel/utils/telemetry/model_diagnostics/gen_ai_attributes.py

@@ -26,6 +26,8 @@
 INPUT_TOKENS = "gen_ai.usage.input_tokens"
 OUTPUT_TOKENS = "gen_ai.usage.output_tokens"
 TOOL_CALL_ID = "gen_ai.tool.call.id"
+TOOL_CALL_ARGUMENTS = "gen_ai.tool.call.arguments"
+TOOL_CALL_RESULT = "gen_ai.tool.call.result"
 TOOL_DESCRIPTION = "gen_ai.tool.description"
 TOOL_NAME = "gen_ai.tool.name"
 ADDRESS = "server.address"