# ------------------------------------
# Copyright (c) Microsoft Corporation.
# Licensed under the MIT License.
# ------------------------------------
import copy
from enum import Enum
import functools
import json
import importlib
import logging
import os
from time import time_ns
from typing import Any, Callable, Dict, Iterator, List, Optional, Tuple, Union
from urllib.parse import urlparse
# pylint: disable = no-name-in-module
from azure.core import CaseInsensitiveEnumMeta # type: ignore
from azure.core.settings import settings
from . import models as _models
try:
# pylint: disable = no-name-in-module
from azure.core.tracing import AbstractSpan, SpanKind # type: ignore
from opentelemetry.trace import StatusCode, Span
_tracing_library_available = True
except ModuleNotFoundError:
_tracing_library_available = False
__all__ = [
"AIInferenceInstrumentor",
]
_inference_traces_enabled: bool = False
_trace_inference_content: bool = False
_INFERENCE_GEN_AI_SYSTEM_NAME = "az.ai.inference"
class TraceType(str, Enum, metaclass=CaseInsensitiveEnumMeta): # pylint: disable=C4747
"""An enumeration class to represent different types of traces."""
INFERENCE = "Inference"
class AIInferenceInstrumentor:
"""
A class for managing the trace instrumentation of AI Inference.
This class allows enabling or disabling tracing for AI Inference.
and provides functionality to check whether instrumentation is active.
"""
def __init__(self):
if not _tracing_library_available:
raise ModuleNotFoundError(
"Azure Core Tracing Opentelemetry is not installed. "
"Please install it using 'pip install azure-core-tracing-opentelemetry'"
)
# In the future we could support different versions from the same library
# and have a parameter that specifies the version to use.
self._impl = _AIInferenceInstrumentorPreview()
def instrument(self, enable_content_recording: Optional[bool] = None) -> None:
"""
Enable trace instrumentation for AI Inference.
:param enable_content_recording: Whether content recording is enabled as part
of the traces or not. Content in this context refers to chat message content
and function call tool related function names, function parameter names and
values. True will enable content recording, False will disable it. If no value
s provided, then the value read from environment variable
AZURE_TRACING_GEN_AI_CONTENT_RECORDING_ENABLED is used. If the environment variable
is not found, then the value will default to False. Please note that successive calls
to instrument will always apply the content recording value provided with the most
recent call to instrument (including applying the environment variable if no value is
provided and defaulting to false if the environment variable is not found), even if
instrument was already previously called without uninstrument being called in between
the instrument calls.
:type enable_content_recording: bool, optional
"""
self._impl.instrument(enable_content_recording=enable_content_recording)
def uninstrument(self) -> None:
"""
Disable trace instrumentation for AI Inference.
Raises:
RuntimeError: If instrumentation is not currently enabled.
This method removes any active instrumentation, stopping the tracing
of AI Inference.
"""
self._impl.uninstrument()
def is_instrumented(self) -> bool:
"""
Check if trace instrumentation for AI Inference is currently enabled.
:return: True if instrumentation is active, False otherwise.
:rtype: bool
"""
return self._impl.is_instrumented()
def is_content_recording_enabled(self) -> bool:
"""
This function gets the content recording value.
:return: A bool value indicating whether content recording is enabled.
:rtype: bool
"""
return self._impl.is_content_recording_enabled()
class _AIInferenceInstrumentorPreview:
"""
A class for managing the trace instrumentation of AI Inference.
This class allows enabling or disabling tracing for AI Inference.
and provides functionality to check whether instrumentation is active.
"""
def _str_to_bool(self, s):
if s is None:
return False
return str(s).lower() == "true"
def instrument(self, enable_content_recording: Optional[bool] = None):
"""
Enable trace instrumentation for AI Inference.
:param enable_content_recording: Whether content recording is enabled as part
of the traces or not. Content in this context refers to chat message content
and function call tool related function names, function parameter names and
values. True will enable content recording, False will disable it. If no value
is provided, then the value read from environment variable
AZURE_TRACING_GEN_AI_CONTENT_RECORDING_ENABLED is used. If the environment variable
is not found, then the value will default to False.
:type enable_content_recording: bool, optional
"""
if enable_content_recording is None:
var_value = os.environ.get("AZURE_TRACING_GEN_AI_CONTENT_RECORDING_ENABLED")
enable_content_recording = self._str_to_bool(var_value)
if not self.is_instrumented():
self._instrument_inference(enable_content_recording)
else:
self._set_content_recording_enabled(enable_content_recording=enable_content_recording)
def uninstrument(self):
"""
Disable trace instrumentation for AI Inference.
This method removes any active instrumentation, stopping the tracing
of AI Inference.
"""
if self.is_instrumented():
self._uninstrument_inference()
def is_instrumented(self):
"""
Check if trace instrumentation for AI Inference is currently enabled.
:return: True if instrumentation is active, False otherwise.
:rtype: bool
"""
return self._is_instrumented()
def set_content_recording_enabled(self, enable_content_recording: bool = False) -> None:
"""This function sets the content recording value.
:param enable_content_recording: Indicates whether tracing of message content should be enabled.
This also controls whether function call tool function names,
parameter names and parameter values are traced.
:type enable_content_recording: bool
"""
self._set_content_recording_enabled(enable_content_recording=enable_content_recording)
def is_content_recording_enabled(self) -> bool:
"""This function gets the content recording value.
:return: A bool value indicating whether content tracing is enabled.
:rtype bool
"""
return self._is_content_recording_enabled()
def _set_attributes(self, span: "AbstractSpan", *attrs: Tuple[str, Any]) -> None:
for attr in attrs:
key, value = attr
if value is not None:
span.add_attribute(key, value)
def _add_request_chat_message_events(self, span: "AbstractSpan", **kwargs: Any) -> int:
timestamp = 0
for message in kwargs.get("messages", []):
try:
message = message.as_dict()
except AttributeError:
pass
if message.get("role"):
timestamp = self._record_event(
span,
f"gen_ai.{message.get('role')}.message",
{
"gen_ai.system": _INFERENCE_GEN_AI_SYSTEM_NAME,
"gen_ai.event.content": json.dumps(message),
},
timestamp,
)
return timestamp
def _parse_url(self, url):
parsed = urlparse(url)
server_address = parsed.hostname
port = parsed.port
return server_address, port
def _add_request_chat_attributes(self, span: "AbstractSpan", *args: Any, **kwargs: Any) -> None:
client = args[0]
endpoint = client._config.endpoint # pylint: disable=protected-access
server_address, port = self._parse_url(endpoint)
model = "chat"
if kwargs.get("model") is not None:
model_value = kwargs.get("model")
if model_value is not None:
model = model_value
self._set_attributes(
span,
("gen_ai.operation.name", "chat"),
("gen_ai.system", _INFERENCE_GEN_AI_SYSTEM_NAME),
("gen_ai.request.model", model),
("gen_ai.request.max_tokens", kwargs.get("max_tokens")),
("gen_ai.request.temperature", kwargs.get("temperature")),
("gen_ai.request.top_p", kwargs.get("top_p")),
("server.address", server_address),
)
if port is not None and port != 443:
span.add_attribute("server.port", port)
def _remove_function_call_names_and_arguments(self, tool_calls: list) -> list:
tool_calls_copy = copy.deepcopy(tool_calls)
for tool_call in tool_calls_copy:
if "function" in tool_call:
if "name" in tool_call["function"]:
del tool_call["function"]["name"]
if "arguments" in tool_call["function"]:
del tool_call["function"]["arguments"]
if not tool_call["function"]:
del tool_call["function"]
return tool_calls_copy
def _get_finish_reasons(self, result) -> Optional[List[str]]:
if hasattr(result, "choices") and result.choices:
finish_reasons: List[str] = []
for choice in result.choices:
finish_reason = getattr(choice, "finish_reason", None)
if finish_reason is None:
# If finish_reason is None, default to "none"
finish_reasons.append("none")
elif hasattr(finish_reason, "value"):
# If finish_reason has a 'value' attribute (i.e., it's an enum), use it
finish_reasons.append(finish_reason.value)
elif isinstance(finish_reason, str):
# If finish_reason is a string, use it directly
finish_reasons.append(finish_reason)
else:
# Default to "none"
finish_reasons.append("none")
return finish_reasons
return None
def _get_finish_reason_for_choice(self, choice):
finish_reason = getattr(choice, "finish_reason", None)
if finish_reason is not None:
return finish_reason.value
return "none"
def _add_response_chat_message_events(
self, span: "AbstractSpan", result: _models.ChatCompletions, last_event_timestamp_ns: int
) -> None:
for choice in result.choices:
attributes = {}
if _trace_inference_content:
full_response: Dict[str, Any] = {
"message": {"content": choice.message.content},
"finish_reason": self._get_finish_reason_for_choice(choice),
"index": choice.index,
}
if choice.message.tool_calls:
full_response["message"]["tool_calls"] = [tool.as_dict() for tool in choice.message.tool_calls]
attributes = {
"gen_ai.system": _INFERENCE_GEN_AI_SYSTEM_NAME,
"gen_ai.event.content": json.dumps(full_response),
}
else:
response: Dict[str, Any] = {
"finish_reason": self._get_finish_reason_for_choice(choice),
"index": choice.index,
}
if choice.message.tool_calls:
response["message"] = {}
tool_calls_function_names_and_arguments_removed = self._remove_function_call_names_and_arguments(
choice.message.tool_calls
)
response["message"]["tool_calls"] = [
tool.as_dict() for tool in tool_calls_function_names_and_arguments_removed
]
attributes = {
"gen_ai.system": _INFERENCE_GEN_AI_SYSTEM_NAME,
"gen_ai.event.content": json.dumps(response),
}
last_event_timestamp_ns = self._record_event(span, "gen_ai.choice", attributes, last_event_timestamp_ns)
def _add_response_chat_attributes(
self,
span: "AbstractSpan",
result: Union[_models.ChatCompletions, _models.StreamingChatCompletionsUpdate],
) -> None:
self._set_attributes(
span,
("gen_ai.response.id", result.id),
("gen_ai.response.model", result.model),
(
"gen_ai.usage.input_tokens",
(result.usage.prompt_tokens if hasattr(result, "usage") and result.usage else None),
),
(
"gen_ai.usage.output_tokens",
(result.usage.completion_tokens if hasattr(result, "usage") and result.usage else None),
),
)
finish_reasons = self._get_finish_reasons(result)
if not finish_reasons is None:
span.add_attribute("gen_ai.response.finish_reasons", finish_reasons) # type: ignore
def _add_request_details(self, span: "AbstractSpan", args: Any, kwargs: Any) -> int:
self._add_request_chat_attributes(span, *args, **kwargs)
if _trace_inference_content:
return self._add_request_chat_message_events(span, **kwargs)
return 0
def _add_response_details(self, span: "AbstractSpan", result: object, last_event_timestamp_ns: int) -> None:
if isinstance(result, _models.ChatCompletions):
self._add_response_chat_attributes(span, result)
self._add_response_chat_message_events(span, result, last_event_timestamp_ns)
# TODO add more models here
def _accumulate_response(self, item, accumulate: Dict[str, Any]) -> None:
if item.finish_reason:
accumulate["finish_reason"] = item.finish_reason
if item.index:
accumulate["index"] = item.index
if item.delta.content:
accumulate.setdefault("message", {})
accumulate["message"].setdefault("content", "")
accumulate["message"]["content"] += item.delta.content
if item.delta.tool_calls:
accumulate.setdefault("message", {})
accumulate["message"].setdefault("tool_calls", [])
if item.delta.tool_calls is not None:
for tool_call in item.delta.tool_calls:
if tool_call.id:
accumulate["message"]["tool_calls"].append(
{
"id": tool_call.id,
"type": "",
"function": {"name": "", "arguments": ""},
}
)
if tool_call.function:
accumulate["message"]["tool_calls"][-1]["type"] = "function"
if tool_call.function and tool_call.function.name:
accumulate["message"]["tool_calls"][-1]["function"]["name"] = tool_call.function.name
if tool_call.function and tool_call.function.arguments:
accumulate["message"]["tool_calls"][-1]["function"]["arguments"] += tool_call.function.arguments
def _accumulate_async_streaming_response(self, item, accumulate: Dict[str, Any]) -> None:
if not "choices" in item:
return
if "finish_reason" in item["choices"][0] and item["choices"][0]["finish_reason"]:
accumulate["finish_reason"] = item["choices"][0]["finish_reason"]
if "index" in item["choices"][0] and item["choices"][0]["index"]:
accumulate["index"] = item["choices"][0]["index"]
if not "delta" in item["choices"][0]:
return
if "content" in item["choices"][0]["delta"] and item["choices"][0]["delta"]["content"]:
accumulate.setdefault("message", {})
accumulate["message"].setdefault("content", "")
accumulate["message"]["content"] += item["choices"][0]["delta"]["content"]
if "tool_calls" in item["choices"][0]["delta"] and item["choices"][0]["delta"]["tool_calls"]:
accumulate.setdefault("message", {})
accumulate["message"].setdefault("tool_calls", [])
if item["choices"][0]["delta"]["tool_calls"] is not None:
for tool_call in item["choices"][0]["delta"]["tool_calls"]:
if tool_call.id:
accumulate["message"]["tool_calls"].append(
{
"id": tool_call.id,
"type": "",
"function": {"name": "", "arguments": ""},
}
)
if tool_call.function:
accumulate["message"]["tool_calls"][-1]["type"] = "function"
if tool_call.function and tool_call.function.name:
accumulate["message"]["tool_calls"][-1]["function"]["name"] = tool_call.function.name
if tool_call.function and tool_call.function.arguments:
accumulate["message"]["tool_calls"][-1]["function"]["arguments"] += tool_call.function.arguments
def _wrapped_stream(
self, stream_obj: _models.StreamingChatCompletions, span: "AbstractSpan", previous_event_timestamp: int
) -> _models.StreamingChatCompletions:
class StreamWrapper(_models.StreamingChatCompletions):
def __init__(self, stream_obj, instrumentor):
super().__init__(stream_obj._response)
self._instrumentor = instrumentor
def __iter__( # pyright: ignore [reportIncompatibleMethodOverride]
self,
) -> Iterator[_models.StreamingChatCompletionsUpdate]:
accumulate: Dict[str, Any] = {}
try:
chunk = None
for chunk in stream_obj:
for item in chunk.choices:
self._instrumentor._accumulate_response(item, accumulate)
yield chunk
if chunk is not None:
self._instrumentor._add_response_chat_attributes(span, chunk)
except Exception as exc:
# Set the span status to error
if isinstance(span.span_instance, Span): # pyright: ignore [reportPossiblyUnboundVariable]
span.span_instance.set_status(
StatusCode.ERROR, # pyright: ignore [reportPossiblyUnboundVariable]
description=str(exc),
)
module = exc.__module__ if hasattr(exc, "__module__") and exc.__module__ != "builtins" else ""
error_type = f"{module}.{type(exc).__name__}" if module else type(exc).__name__
self._instrumentor._set_attributes(span, ("error.type", error_type))
raise
finally:
if stream_obj._done is False:
if accumulate.get("finish_reason") is None:
accumulate["finish_reason"] = "error"
else:
# Only one choice expected with streaming
accumulate["index"] = 0
# Delete message if content tracing is not enabled
if not _trace_inference_content:
if "message" in accumulate:
if "content" in accumulate["message"]:
del accumulate["message"]["content"]
if not accumulate["message"]:
del accumulate["message"]
if "message" in accumulate:
if "tool_calls" in accumulate["message"]:
tool_calls_function_names_and_arguments_removed = (
self._instrumentor._remove_function_call_names_and_arguments(
accumulate["message"]["tool_calls"]
)
)
accumulate["message"]["tool_calls"] = list(
tool_calls_function_names_and_arguments_removed
)
attributes = {
"gen_ai.system": _INFERENCE_GEN_AI_SYSTEM_NAME,
"gen_ai.event.content": json.dumps(accumulate),
}
self._instrumentor._record_event(span, "gen_ai.choice", attributes, previous_event_timestamp)
span.finish()
return StreamWrapper(stream_obj, self)
def _async_wrapped_stream(
self, stream_obj: _models.AsyncStreamingChatCompletions, span: "AbstractSpan", last_event_timestamp_ns: int
) -> _models.AsyncStreamingChatCompletions:
class AsyncStreamWrapper(_models.AsyncStreamingChatCompletions):
def __init__(self, stream_obj, instrumentor, span, last_event_timestamp_ns):
super().__init__(stream_obj._response)
self._instrumentor = instrumentor
self._accumulate: Dict[str, Any] = {}
self._stream_obj = stream_obj
self.span = span
self._last_result = None
self._last_event_timestamp_ns = last_event_timestamp_ns
async def __anext__(self) -> "_models.StreamingChatCompletionsUpdate":
try:
result = await super().__anext__()
self._instrumentor._accumulate_async_streaming_response( # pylint: disable=protected-access, line-too-long # pyright: ignore [reportFunctionMemberAccess]
result, self._accumulate
)
self._last_result = result
except StopAsyncIteration as exc:
self._trace_stream_content()
raise exc
return result
def _trace_stream_content(self) -> None:
if self._last_result:
self._instrumentor._add_response_chat_attributes( # pylint: disable=protected-access, line-too-long # pyright: ignore [reportFunctionMemberAccess]
span, self._last_result
)
# Only one choice expected with streaming
self._accumulate["index"] = 0
# Delete message if content tracing is not enabled
if not _trace_inference_content:
if "message" in self._accumulate:
if "content" in self._accumulate["message"]:
del self._accumulate["message"]["content"]
if not self._accumulate["message"]:
del self._accumulate["message"]
if "message" in self._accumulate:
if "tool_calls" in self._accumulate["message"]:
tools_no_recording = self._instrumentor._remove_function_call_names_and_arguments( # pylint: disable=protected-access, line-too-long # pyright: ignore [reportFunctionMemberAccess]
self._accumulate["message"]["tool_calls"]
)
self._accumulate["message"]["tool_calls"] = list(tools_no_recording)
attributes = {
"gen_ai.system": _INFERENCE_GEN_AI_SYSTEM_NAME,
"gen_ai.event.content": json.dumps(self._accumulate),
}
self._last_event_timestamp_ns = self._instrumentor._record_event( # pylint: disable=protected-access, line-too-long # pyright: ignore [reportFunctionMemberAccess]
span, "gen_ai.choice", attributes, self._last_event_timestamp_ns
)
span.finish()
async_stream_wrapper = AsyncStreamWrapper(stream_obj, self, span, last_event_timestamp_ns)
return async_stream_wrapper
def _record_event(
self, span: "AbstractSpan", name: str, attributes: Dict[str, Any], last_event_timestamp_ns: int
) -> int:
timestamp = time_ns()
# we're recording multiple events, some of them are emitted within (hundreds of) nanoseconds of each other.
# time.time_ns resolution is not high enough on windows to guarantee unique timestamps for each message.
# Also Azure Monitor truncates resolution to microseconds and some other backends truncate to milliseconds.
#
# But we need to give users a way to restore event order, so we're incrementing the timestamp
# by 1 microsecond for each message.
#
# This is a workaround, we'll find a generic and better solution - see
# https://github.com/open-telemetry/semantic-conventions/issues/1701
if last_event_timestamp_ns > 0 and timestamp <= (last_event_timestamp_ns + 1000):
timestamp = last_event_timestamp_ns + 1000
span.span_instance.add_event(name=name, attributes=attributes, timestamp=timestamp)
return timestamp
def _trace_sync_function(
self,
function: Callable,
*,
_args_to_ignore: Optional[List[str]] = None,
_trace_type=TraceType.INFERENCE,
_name: Optional[str] = None,
) -> Callable:
"""
Decorator that adds tracing to a synchronous function.
:param function: The function to be traced.
:type function: Callable
:param args_to_ignore: A list of argument names to be ignored in the trace.
Defaults to None.
:type: args_to_ignore: [List[str]], optional
:param trace_type: The type of the trace. Defaults to TraceType.INFERENCE.
:type trace_type: TraceType, optional
:param name: The name of the trace, will set to func name if not provided.
:type name: str, optional
:return: The traced function.
:rtype: Callable
"""
@functools.wraps(function)
def inner(*args, **kwargs):
span_impl_type = settings.tracing_implementation()
if span_impl_type is None:
return function(*args, **kwargs)
class_function_name = function.__qualname__
if class_function_name.startswith("ChatCompletionsClient.complete"):
if kwargs.get("model") is None:
span_name = "chat"
else:
model = kwargs.get("model")
span_name = f"chat {model}"
span = span_impl_type(
name=span_name,
kind=SpanKind.CLIENT, # pyright: ignore [reportPossiblyUnboundVariable]
)
try:
# tracing events not supported in azure-core-tracing-opentelemetry
# so need to access the span instance directly
with span_impl_type.change_context(span.span_instance):
last_event_timestamp_ns = self._add_request_details(span, args, kwargs)
result = function(*args, **kwargs)
if kwargs.get("stream") is True:
return self._wrapped_stream(result, span, last_event_timestamp_ns)
self._add_response_details(span, result, last_event_timestamp_ns)
except Exception as exc:
# Set the span status to error
if isinstance(span.span_instance, Span): # pyright: ignore [reportPossiblyUnboundVariable]
span.span_instance.set_status(
StatusCode.ERROR, # pyright: ignore [reportPossiblyUnboundVariable]
description=str(exc),
)
module = getattr(exc, "__module__", "")
module = module if module != "builtins" else ""
error_type = f"{module}.{type(exc).__name__}" if module else type(exc).__name__
self._set_attributes(span, ("error.type", error_type))
span.finish()
raise
span.finish()
return result
# Handle the default case (if the function name does not match)
return None # Ensure all paths return
return inner
def _trace_async_function(
self,
function: Callable,
*,
_args_to_ignore: Optional[List[str]] = None,
_trace_type=TraceType.INFERENCE,
_name: Optional[str] = None,
) -> Callable:
"""
Decorator that adds tracing to an asynchronous function.
:param function: The function to be traced.
:type function: Callable
:param args_to_ignore: A list of argument names to be ignored in the trace.
Defaults to None.
:type: args_to_ignore: [List[str]], optional
:param trace_type: The type of the trace. Defaults to TraceType.INFERENCE.
:type trace_type: TraceType, optional
:param name: The name of the trace, will set to func name if not provided.
:type name: str, optional
:return: The traced function.
:rtype: Callable
"""
@functools.wraps(function)
async def inner(*args, **kwargs):
span_impl_type = settings.tracing_implementation()
if span_impl_type is None:
return await function(*args, **kwargs)
class_function_name = function.__qualname__
if class_function_name.startswith("ChatCompletionsClient.complete"):
if kwargs.get("model") is None:
span_name = "chat"
else:
model = kwargs.get("model")
span_name = f"chat {model}"
span = span_impl_type(
name=span_name,
kind=SpanKind.CLIENT, # pyright: ignore [reportPossiblyUnboundVariable]
)
try:
# tracing events not supported in azure-core-tracing-opentelemetry
# so need to access the span instance directly
with span_impl_type.change_context(span.span_instance):
last_event_timestamp_ns = self._add_request_details(span, args, kwargs)
result = await function(*args, **kwargs)
if kwargs.get("stream") is True:
return self._async_wrapped_stream(result, span, last_event_timestamp_ns)
self._add_response_details(span, result, last_event_timestamp_ns)
except Exception as exc:
# Set the span status to error
if isinstance(span.span_instance, Span): # pyright: ignore [reportPossiblyUnboundVariable]
span.span_instance.set_status(
StatusCode.ERROR, # pyright: ignore [reportPossiblyUnboundVariable]
description=str(exc),
)
module = getattr(exc, "__module__", "")
module = module if module != "builtins" else ""
error_type = f"{module}.{type(exc).__name__}" if module else type(exc).__name__
self._set_attributes(span, ("error.type", error_type))
span.finish()
raise
span.finish()
return result
# Handle the default case (if the function name does not match)
return None # Ensure all paths return
return inner
def _inject_async(self, f, _trace_type, _name):
wrapper_fun = self._trace_async_function(f)
wrapper_fun._original = f # pylint: disable=protected-access # pyright: ignore [reportFunctionMemberAccess]
return wrapper_fun
def _inject_sync(self, f, _trace_type, _name):
wrapper_fun = self._trace_sync_function(f)
wrapper_fun._original = f # pylint: disable=protected-access # pyright: ignore [reportFunctionMemberAccess]
return wrapper_fun
def _inference_apis(self):
sync_apis = (
(
"azure.ai.inference",
"ChatCompletionsClient",
"complete",
TraceType.INFERENCE,
"inference_chat_completions_complete",
),
)
async_apis = (
(
"azure.ai.inference.aio",
"ChatCompletionsClient",
"complete",
TraceType.INFERENCE,
"inference_chat_completions_complete",
),
)
return sync_apis, async_apis
def _inference_api_list(self):
sync_apis, async_apis = self._inference_apis()
yield sync_apis, self._inject_sync
yield async_apis, self._inject_async
def _generate_api_and_injector(self, apis):
for api, injector in apis:
for module_name, class_name, method_name, trace_type, name in api:
try:
module = importlib.import_module(module_name)
api = getattr(module, class_name)
if hasattr(api, method_name):
yield api, method_name, trace_type, injector, name
except AttributeError as e:
# Log the attribute exception with the missing class information
logging.warning(
"AttributeError: The module '%s' does not have the class '%s'. %s",
module_name,
class_name,
str(e),
)
except Exception as e: # pylint: disable=broad-except
# Log other exceptions as a warning, as we're not sure what they might be
logging.warning("An unexpected error occurred: '%s'", str(e))
def _available_inference_apis_and_injectors(self):
"""
Generates a sequence of tuples containing Inference API classes, method names, and
corresponding injector functions.
:return: A generator yielding tuples.
:rtype: tuple
"""
yield from self._generate_api_and_injector(self._inference_api_list())
def _instrument_inference(self, enable_content_tracing: bool = False):
"""This function modifies the methods of the Inference API classes to
inject logic before calling the original methods.
The original methods are stored as _original attributes of the methods.
:param enable_content_tracing: Indicates whether tracing of message content should be enabled.
This also controls whether function call tool function names,
parameter names and parameter values are traced.
:type enable_content_tracing: bool
"""
# pylint: disable=W0603
global _inference_traces_enabled
global _trace_inference_content
if _inference_traces_enabled:
raise RuntimeError("Traces already started for azure.ai.inference")
_inference_traces_enabled = True
_trace_inference_content = enable_content_tracing
for (
api,
method,
trace_type,
injector,
name,
) in self._available_inference_apis_and_injectors():
# Check if the method of the api class has already been modified
if not hasattr(getattr(api, method), "_original"):
setattr(api, method, injector(getattr(api, method), trace_type, name))
def _uninstrument_inference(self):
"""This function restores the original methods of the Inference API classes
by assigning them back from the _original attributes of the modified methods.
"""
# pylint: disable=W0603
global _inference_traces_enabled
global _trace_inference_content
_trace_inference_content = False
for api, method, _, _, _ in self._available_inference_apis_and_injectors():
if hasattr(getattr(api, method), "_original"):
setattr(api, method, getattr(getattr(api, method), "_original"))
_inference_traces_enabled = False
def _is_instrumented(self):
"""This function returns True if Inference libary has already been instrumented
for tracing and False if it has not been instrumented.
:return: A value indicating whether the Inference library is currently instrumented or not.
:rtype: bool
"""
return _inference_traces_enabled
def _set_content_recording_enabled(self, enable_content_recording: bool = False) -> None:
"""This function sets the content recording value.
:param enable_content_recording: Indicates whether tracing of message content should be enabled.
This also controls whether function call tool function names,
parameter names and parameter values are traced.
:type enable_content_recording: bool
"""
global _trace_inference_content # pylint: disable=W0603
_trace_inference_content = enable_content_recording
def _is_content_recording_enabled(self) -> bool:
"""This function gets the content recording value.
:return: A bool value indicating whether content tracing is enabled.
:rtype bool
"""
return _trace_inference_content
