Skip to content

Same agent Smaller bill. The tools and infrastructure to get more work from your AI. Fewer tokens. Less compute. Lower cost.

[Feat]: Use Events API In Azure AI Inference OpenTelemetry Instrumentation (#662)

Patcher117fe1132760
diff --git a/sdk/python/pyproject.toml b/sdk/python/pyproject.toml
index a4ff4b30..e9ddd19d 100644
--- a/sdk/python/pyproject.toml
+++ b/sdk/python/pyproject.toml
@@ -1,6 +1,6 @@
[tool.poetry]
name = "openlit"
-version = "1.33.17"
+version = "1.33.18"
description = "OpenTelemetry-native Auto instrumentation library for monitoring LLM Applications and GPUs, facilitating the integration of observability into your GenAI-driven projects"
authors = ["OpenLIT"]
license = "Apache-2.0"
diff --git a/sdk/python/src/openlit/instrumentation/azure_ai_inference/__init__.py b/sdk/python/src/openlit/instrumentation/azure_ai_inference/__init__.py
index d946b80d..b4ea879c 100644
--- a/sdk/python/src/openlit/instrumentation/azure_ai_inference/__init__.py
+++ b/sdk/python/src/openlit/instrumentation/azure_ai_inference/__init__.py
@@ -4,13 +4,11 @@ from typing import Collection
import importlib.metadata
from opentelemetry.instrumentation.instrumentor import BaseInstrumentor
from wrapt import wrap_function_wrapper
-
from openlit.instrumentation.azure_ai_inference.azure_ai_inference import (
- complete, embedding
+ complete
)
-
from openlit.instrumentation.azure_ai_inference.async_azure_ai_inference import (
- async_complete, async_embedding
+ async_complete
)
_instruments = ('azure-ai-inference >= 1.0.0b4',)
@@ -27,6 +25,7 @@ class AzureAIInferenceInstrumentor(BaseInstrumentor):
application_name = kwargs.get('application_name', 'default')
environment = kwargs.get('environment', 'default')
tracer = kwargs.get('tracer')
+ event_provider = kwargs.get('event_provider')
metrics = kwargs.get('metrics_dict')
pricing_info = kwargs.get('pricing_info', {})
capture_message_content = kwargs.get('capture_message_content', False)
@@ -38,15 +37,7 @@ class AzureAIInferenceInstrumentor(BaseInstrumentor):
'azure.ai.inference',
'ChatCompletionsClient.complete',
complete(version, environment, application_name,
- tracer, pricing_info, capture_message_content, metrics, disable_metrics),
- )
-
- # sync embedding
- wrap_function_wrapper(
- 'azure.ai.inference',
- 'EmbeddingsClient.embed',
- embedding(version, environment, application_name,
- tracer, pricing_info, capture_message_content, metrics, disable_metrics),
+ tracer, event_provider, pricing_info, capture_message_content, metrics, disable_metrics),
)
# async generate
@@ -54,15 +45,7 @@ class AzureAIInferenceInstrumentor(BaseInstrumentor):
'azure.ai.inference.aio',
'ChatCompletionsClient.complete',
async_complete(version, environment, application_name,
- tracer, pricing_info, capture_message_content, metrics, disable_metrics),
- )
-
- # async embedding
- wrap_function_wrapper(
- 'azure.ai.inference.aio',
- 'EmbeddingsClient.embed',
- async_embedding(version, environment, application_name,
- tracer, pricing_info, capture_message_content, metrics, disable_metrics),
+ tracer, event_provider, pricing_info, capture_message_content, metrics, disable_metrics),
)
def _uninstrument(self, **kwargs):
diff --git a/sdk/python/src/openlit/instrumentation/azure_ai_inference/async_azure_ai_inference.py b/sdk/python/src/openlit/instrumentation/azure_ai_inference/async_azure_ai_inference.py
index 429473ed..522595f1 100644
--- a/sdk/python/src/openlit/instrumentation/azure_ai_inference/async_azure_ai_inference.py
+++ b/sdk/python/src/openlit/instrumentation/azure_ai_inference/async_azure_ai_inference.py
@@ -4,18 +4,15 @@ Module for monitoring Azure AI Inference API calls.
import logging
import time
-from opentelemetry.trace import SpanKind, Status, StatusCode
-from opentelemetry.sdk.resources import SERVICE_NAME, TELEMETRY_SDK_NAME, DEPLOYMENT_ENVIRONMENT
+from opentelemetry.trace import SpanKind
from openlit.__helpers import (
- get_chat_model_cost,
- get_embed_model_cost,
handle_exception,
- response_as_dict,
- calculate_ttft,
- calculate_tbt,
- create_metrics_attributes,
set_server_address_and_port,
- general_tokens
+)
+from openlit.instrumentation.azure_ai_inference.utils import (
+ process_chunk,
+ process_chat_response,
+ process_streaming_chat_response,
)
from openlit.semcov import SemanticConvetion
@@ -23,37 +20,21 @@ from openlit.semcov import SemanticConvetion
logger = logging.getLogger(__name__)
def async_complete(version, environment, application_name,
- tracer, pricing_info, capture_message_content, metrics, disable_metrics):
+ tracer, event_provider, pricing_info, capture_message_content, metrics, disable_metrics):
"""
- Generates a telemetry wrapper for chat to collect metrics.
-
- Args:
- version: Version of the monitoring package.
- environment: Deployment environment (e.g., production, staging).
- application_name: Name of the application using the Azure AI Inference API.
- tracer: OpenTelemetry tracer for creating spans.
- pricing_info: Information used for calculating the cost of Azure AI Inference usage.
- capture_message_content: Flag indicating whether to trace the actual content.
-
- Returns:
- A function that wraps the chat method to add telemetry.
+ Generates a telemetry wrapper for GenAI function call
"""
class TracedAsyncStream:
"""
- Wrapper for streaming responses to collect metrics and trace data.
- Wraps the response to collect message IDs and aggregated response.
-
- This class implements the '__aiter__' and '__anext__' methods that
- handle asynchronous streaming responses.
-
- This class also implements '__aenter__' and '__aexit__' methods that
- handle asynchronous context management protocol.
+ Wrapper for streaming responses to collect telemetry.
"""
+
def __init__(
self,
wrapped,
span,
+ span_name,
kwargs,
server_address,
server_port,
@@ -61,12 +42,13 @@ def async_complete(version, environment, application_name,
):
self.__wrapped__ = wrapped
self._span = span
- # Placeholder for aggregating streaming response
+ self._span_name = span_name
self._llmresponse = ""
self._response_id = ""
self._response_model = ""
self._finish_reason = ""
- self._system_fingerprint = ""
+ self._input_tokens = 0
+ self._output_tokens = 0
self._args = args
self._kwargs = kwargs
@@ -95,197 +77,33 @@ def async_complete(version, environment, application_name,
async def __anext__(self):
try:
chunk = await self.__wrapped__.__anext__()
- end_time = time.time()
- # Record the timestamp for the current chunk
- self._timestamps.append(end_time)
-
- if len(self._timestamps) == 1:
- # Calculate time to first chunk
- self._ttft = calculate_ttft(self._timestamps, self._start_time)
-
- chunked = response_as_dict(chunk)
- # Collect message IDs and aggregated response from events
- if (len(chunked.get('choices')) > 0 and ('delta' in chunked.get('choices')[0] and
- 'content' in chunked.get('choices')[0].get('delta'))):
-
- content = chunked.get('choices')[0].get('delta').get('content')
- if content:
- self._llmresponse += content
- self._response_id = chunked.get('id')
- self._response_model = chunked.get('model')
- self._finish_reason = chunked.get('choices')[0].get('finish_reason')
- self._system_fingerprint = chunked.get('system_fingerprint')
+ process_chunk(self, chunk)
return chunk
except StopAsyncIteration:
- # Handling exception ensure observability without disrupting operation
try:
- self._end_time = time.time()
- if len(self._timestamps) > 1:
- self._tbt = calculate_tbt(self._timestamps)
-
- # Format 'messages' into a single string
- message_prompt = self._kwargs.get("messages", "")
- formatted_messages = []
- for message in message_prompt:
- role = message["role"]
- content = message["content"]
-
- if isinstance(content, list):
- content_str_list = []
- for item in content:
- if item["type"] == "text":
- content_str_list.append(f'text: {item["text"]}')
- elif (item["type"] == "image_url" and
- not item["image_url"]["url"].startswith("data:")):
- content_str_list.append(f'image_url: {item["image_url"]["url"]}')
- content_str = ", ".join(content_str_list)
- formatted_messages.append(f"{role}: {content_str}")
- else:
- formatted_messages.append(f"{role}: {content}")
- prompt = "\n".join(formatted_messages)
-
- request_model = self._kwargs.get("model", "gpt-4o")
-
- # Calculate tokens using input prompt and aggregated response
- input_tokens = general_tokens(prompt)
- output_tokens = general_tokens(self._llmresponse)
-
- # Calculate cost of the operation
- cost = get_chat_model_cost(request_model,
- pricing_info, input_tokens,
- output_tokens)
-
- # Set Span attributes (OTel Semconv)
- self._span.set_attribute(TELEMETRY_SDK_NAME, "openlit")
- self._span.set_attribute(SemanticConvetion.GEN_AI_OPERATION,
- SemanticConvetion.GEN_AI_OPERATION_TYPE_CHAT)
- self._span.set_attribute(SemanticConvetion.GEN_AI_SYSTEM,
- SemanticConvetion.GEN_AI_SYSTEM_AZURE_AI_INFERENCE)
- self._span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_MODEL,
- request_model)
- self._span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_SEED,
- self._kwargs.get("seed", ""))
- self._span.set_attribute(SemanticConvetion.SERVER_PORT,
- self._server_port)
- self._span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_FREQUENCY_PENALTY,
- self._kwargs.get("frequency_penalty", 0.0))
- self._span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_MAX_TOKENS,
- self._kwargs.get("max_tokens", -1))
- self._span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_PRESENCE_PENALTY,
- self._kwargs.get("presence_penalty", 0.0))
- self._span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_STOP_SEQUENCES,
- self._kwargs.get("stop", []))
- self._span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_TEMPERATURE,
- self._kwargs.get("temperature", 1.0))
- self._span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_TOP_P,
- self._kwargs.get("top_p", 1.0))
- self._span.set_attribute(SemanticConvetion.GEN_AI_RESPONSE_FINISH_REASON,
- [self._finish_reason])
- self._span.set_attribute(SemanticConvetion.GEN_AI_RESPONSE_ID,
- self._response_id)
- self._span.set_attribute(SemanticConvetion.GEN_AI_RESPONSE_MODEL,
- self._response_model)
- self._span.set_attribute(SemanticConvetion.GEN_AI_USAGE_INPUT_TOKENS,
- input_tokens)
- self._span.set_attribute(SemanticConvetion.GEN_AI_USAGE_OUTPUT_TOKENS,
- output_tokens)
- self._span.set_attribute(SemanticConvetion.SERVER_ADDRESS,
- self._server_address)
- self._span.set_attribute(SemanticConvetion.GEN_AI_RESPONSE_SYSTEM_FINGERPRINT,
- self._system_fingerprint)
- if isinstance(self._llmresponse, str):
- self._span.set_attribute(SemanticConvetion.GEN_AI_OUTPUT_TYPE,
- "text")
- else:
- self._span.set_attribute(SemanticConvetion.GEN_AI_OUTPUT_TYPE,
- "json")
-
- # Set Span attributes (Extra)
- self._span.set_attribute(DEPLOYMENT_ENVIRONMENT,
- environment)
- self._span.set_attribute(SERVICE_NAME,
- application_name)
- self._span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_IS_STREAM,
- True)
- self._span.set_attribute(SemanticConvetion.GEN_AI_USAGE_TOTAL_TOKENS,
- input_tokens + output_tokens)
- self._span.set_attribute(SemanticConvetion.GEN_AI_USAGE_COST,
- cost)
- self._span.set_attribute(SemanticConvetion.GEN_AI_SERVER_TBT,
- self._tbt)
- self._span.set_attribute(SemanticConvetion.GEN_AI_SERVER_TTFT,
- self._ttft)
- self._span.set_attribute(SemanticConvetion.GEN_AI_SDK_VERSION,
- version)
- if capture_message_content:
- self._span.add_event(
- name=SemanticConvetion.GEN_AI_CONTENT_PROMPT_EVENT,
- attributes={
- SemanticConvetion.GEN_AI_CONTENT_PROMPT: prompt,
- },
- )
- self._span.add_event(
- name=SemanticConvetion.GEN_AI_CONTENT_COMPLETION_EVENT,
- attributes={
- SemanticConvetion.GEN_AI_CONTENT_COMPLETION: self._llmresponse,
- },
+ with tracer.start_as_current_span(self._span_name, kind= SpanKind.CLIENT) as self._span:
+ process_streaming_chat_response(
+ self,
+ pricing_info=pricing_info,
+ environment=environment,
+ application_name=application_name,
+ metrics=metrics,
+ event_provider=event_provider,
+ capture_message_content=capture_message_content,
+ disable_metrics=disable_metrics,
+ version=version
)
- self._span.set_status(Status(StatusCode.OK))
-
- if disable_metrics is False:
- attributes = create_metrics_attributes(
- service_name=application_name,
- deployment_environment=environment,
- operation=SemanticConvetion.GEN_AI_OPERATION_TYPE_CHAT,
- system=SemanticConvetion.GEN_AI_SYSTEM_AZURE_AI_INFERENCE,
- request_model=request_model,
- server_address=self._server_address,
- server_port=self._server_port,
- response_model=self._response_model,
- )
-
- metrics["genai_client_usage_tokens"].record(
- input_tokens + output_tokens, attributes
- )
- metrics["genai_client_operation_duration"].record(
- self._end_time - self._start_time, attributes
- )
- metrics["genai_server_tbt"].record(
- self._tbt, attributes
- )
- metrics["genai_server_ttft"].record(
- self._ttft, attributes
- )
- metrics["genai_requests"].add(1, attributes)
- metrics["genai_completion_tokens"].add(output_tokens, attributes)
- metrics["genai_prompt_tokens"].add(input_tokens, attributes)
- metrics["genai_cost"].record(cost, attributes)
except Exception as e:
handle_exception(self._span, e)
logger.error("Error in trace creation: %s", e)
- finally:
- self._span.end()
raise
async def wrapper(wrapped, instance, args, kwargs):
"""
- Wraps the 'chat.completions' API call to add telemetry.
-
- This collects metrics such as execution time, cost, and token usage, and handles errors
- gracefully, adding details to the trace for observability.
-
- Args:
- wrapped: The original 'chat.completions' method to be wrapped.
- instance: The instance of the class where the original method is defined.
- args: Positional arguments for the 'chat.completions' method.
- kwargs: Keyword arguments for the 'chat.completions' method.
-
- Returns:
- The response from the original 'chat.completions' method.
+ Wraps the GenAI function call.
"""
- # Check if streaming is enabled for the API call
streaming = kwargs.get("stream", False)
server_address, server_port = set_server_address_and_port(instance, "models.github.ai", 443)
request_model = kwargs.get("model", "gpt-4o")
@@ -294,292 +112,33 @@ def async_complete(version, environment, application_name,
# pylint: disable=no-else-return
if streaming:
- # Special handling for streaming response to accommodate the nature of data flow
awaited_wrapped = await wrapped(*args, **kwargs)
span = tracer.start_span(span_name, kind=SpanKind.CLIENT)
- return TracedAsyncStream(awaited_wrapped, span, kwargs, server_address, server_port)
+ return TracedAsyncStream(awaited_wrapped, span, span_name, kwargs, server_address, server_port)
- # Handling for non-streaming responses
else:
- with tracer.start_as_current_span(span_name, kind= SpanKind.CLIENT) as span:
+ with tracer.start_as_current_span(span_name, kind=SpanKind.CLIENT) as span:
start_time = time.time()
response = await wrapped(*args, **kwargs)
- end_time = time.time()
-
- response_dict = response_as_dict(response)
-
- try:
- # Format 'messages' into a single string
- message_prompt = kwargs.get("messages", "")
- formatted_messages = []
- for message in message_prompt:
- role = message["role"]
- content = message["content"]
-
- if isinstance(content, list):
- content_str = ", ".join(
- f'{item["type"]}: {item["text"] if "text" in item else item["image_url"]}'
- if "type" in item else f'text: {item["text"]}'
- for item in content
- )
- formatted_messages.append(f"{role}: {content_str}")
- else:
- formatted_messages.append(f"{role}: {content}")
- prompt = "\n".join(formatted_messages)
-
- input_tokens = response_dict.get('usage').get('prompt_tokens')
- output_tokens = response_dict.get('usage').get('completion_tokens')
-
- # Calculate cost of the operation
- cost = get_chat_model_cost(request_model,
- pricing_info, input_tokens,
- output_tokens)
-
- # Set base span attribues (OTel Semconv)
- span.set_attribute(TELEMETRY_SDK_NAME, "openlit")
- span.set_attribute(SemanticConvetion.GEN_AI_OPERATION,
- SemanticConvetion.GEN_AI_OPERATION_TYPE_CHAT)
- span.set_attribute(SemanticConvetion.GEN_AI_SYSTEM,
- SemanticConvetion.GEN_AI_SYSTEM_AZURE_AI_INFERENCE)
- span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_MODEL,
- request_model)
- span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_SEED,
- kwargs.get("seed", ""))
- span.set_attribute(SemanticConvetion.SERVER_PORT,
- server_port)
- span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_FREQUENCY_PENALTY,
- kwargs.get("frequency_penalty", 0.0))
- span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_MAX_TOKENS,
- kwargs.get("max_tokens", -1))
- span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_PRESENCE_PENALTY,
- kwargs.get("presence_penalty", 0.0))
- span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_STOP_SEQUENCES,
- kwargs.get("stop", []))
- span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_TEMPERATURE,
- kwargs.get("temperature", 1.0))
- span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_TOP_P,
- kwargs.get("top_p", 1.0))
- span.set_attribute(SemanticConvetion.GEN_AI_RESPONSE_ID,
- response_dict.get("id"))
- span.set_attribute(SemanticConvetion.GEN_AI_RESPONSE_MODEL,
- response_dict.get('model'))
- span.set_attribute(SemanticConvetion.GEN_AI_USAGE_INPUT_TOKENS,
- input_tokens)
- span.set_attribute(SemanticConvetion.GEN_AI_USAGE_OUTPUT_TOKENS,
- output_tokens)
- span.set_attribute(SemanticConvetion.SERVER_ADDRESS,
- server_address)
- span.set_attribute(SemanticConvetion.GEN_AI_RESPONSE_SYSTEM_FINGERPRINT,
- response_dict.get('system_fingerprint'))
-
- # Set base span attribues (Extras)
- span.set_attribute(DEPLOYMENT_ENVIRONMENT,
- environment)
- span.set_attribute(SERVICE_NAME,
- application_name)
- span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_IS_STREAM,
- False)
- span.set_attribute(SemanticConvetion.GEN_AI_USAGE_TOTAL_TOKENS,
- input_tokens + output_tokens)
- span.set_attribute(SemanticConvetion.GEN_AI_USAGE_COST,
- cost)
- span.set_attribute(SemanticConvetion.GEN_AI_SERVER_TTFT,
- end_time - start_time)
- span.set_attribute(SemanticConvetion.GEN_AI_SDK_VERSION,
- version)
- if capture_message_content:
- span.add_event(
- name=SemanticConvetion.GEN_AI_CONTENT_PROMPT_EVENT,
- attributes={
- SemanticConvetion.GEN_AI_CONTENT_PROMPT: prompt,
- },
- )
-
- for i in range(kwargs.get('n',1)):
- span.set_attribute(SemanticConvetion.GEN_AI_RESPONSE_FINISH_REASON,
- [response_dict.get('choices')[i].get('finish_reason')])
- if capture_message_content:
- span.add_event(
- name=SemanticConvetion.GEN_AI_CONTENT_COMPLETION_EVENT,
- attributes={
- # pylint: disable=line-too-long
- SemanticConvetion.GEN_AI_CONTENT_COMPLETION: str(response_dict.get('choices')[i].get('message').get('content')),
- },
- )
- if kwargs.get('tools'):
- span.set_attribute(SemanticConvetion.GEN_AI_TOOL_CALLS,
- str(response_dict.get('choices')[i].get('message').get('tool_calls')))
-
- if isinstance(response_dict.get('choices')[i].get('message').get('content'), str):
- span.set_attribute(SemanticConvetion.GEN_AI_OUTPUT_TYPE,
- "text")
- elif response_dict.get('choices')[i].get('message').get('content') is not None:
- span.set_attribute(SemanticConvetion.GEN_AI_OUTPUT_TYPE,
- "json")
-
- span.set_status(Status(StatusCode.OK))
-
- if disable_metrics is False:
- attributes = create_metrics_attributes(
- service_name=application_name,
- deployment_environment=environment,
- operation=SemanticConvetion.GEN_AI_OPERATION_TYPE_CHAT,
- system=SemanticConvetion.GEN_AI_SYSTEM_AZURE_AI_INFERENCE,
- request_model=request_model,
- server_address=server_address,
- server_port=server_port,
- response_model=response_dict.get('model'),
- )
-
- metrics["genai_client_usage_tokens"].record(
- input_tokens + output_tokens, attributes
- )
- metrics["genai_client_operation_duration"].record(
- end_time - start_time, attributes
- )
- metrics["genai_server_ttft"].record(
- end_time - start_time, attributes
- )
- metrics["genai_requests"].add(1, attributes)
- metrics["genai_completion_tokens"].add(output_tokens, attributes)
- metrics["genai_prompt_tokens"].add(input_tokens, attributes)
- metrics["genai_cost"].record(cost, attributes)
-
- # Return original response
- return response
-
- except Exception as e:
- handle_exception(span, e)
- logger.error("Error in trace creation: %s", e)
-
- # Return original response
- return response
-
- return wrapper
-
-def async_embedding(version, environment, application_name,
- tracer, pricing_info, capture_message_content, metrics, disable_metrics):
- """
- Generates a telemetry wrapper for embeddings to collect metrics.
-
- Args:
- version: Version of the monitoring package.
- environment: Deployment environment (e.g., production, staging).
- application_name: Name of the application using the Azure Inference API.
- tracer: OpenTelemetry tracer for creating spans.
- pricing_info: Information used for calculating the cost of Azure Inference usage.
- capture_message_content: Flag indicating whether to trace the actual content.
-
- Returns:
- A function that wraps the embeddings method to add telemetry.
- """
-
- async def wrapper(wrapped, instance, args, kwargs):
- """
- Wraps the 'embeddings' API call to add telemetry.
-
- This collects metrics such as execution time, cost, and token usage, and handles errors
- gracefully, adding details to the trace for observability.
-
- Args:
- wrapped: The original 'embeddings' method to be wrapped.
- instance: The instance of the class where the original method is defined.
- args: Positional arguments for the 'embeddings' method.
- kwargs: Keyword arguments for the 'embeddings' method.
-
- Returns:
- The response from the original 'embeddings' method.
- """
-
- server_address, server_port = set_server_address_and_port(instance, "models.github.ai", 443)
- request_model = kwargs.get("model", "text-embedding-ada-002")
-
- span_name = f"{SemanticConvetion.GEN_AI_OPERATION_TYPE_EMBEDDING} {request_model}"
-
- with tracer.start_as_current_span(span_name, kind= SpanKind.CLIENT) as span:
- start_time = time.time()
- response = await wrapped(*args, **kwargs)
- end_time = time.time()
-
- response_dict = response_as_dict(response)
- try:
- input_tokens = response_dict.get('usage').get('prompt_tokens')
-
- # Calculate cost of the operation
- cost = get_embed_model_cost(request_model,
- pricing_info, input_tokens)
-
- # Set Span attributes (OTel Semconv)
- span.set_attribute(TELEMETRY_SDK_NAME, "openlit")
- span.set_attribute(SemanticConvetion.GEN_AI_OPERATION,
- SemanticConvetion.GEN_AI_OPERATION_TYPE_EMBEDDING)
- span.set_attribute(SemanticConvetion.GEN_AI_SYSTEM,
- SemanticConvetion.GEN_AI_SYSTEM_AZURE_AI_INFERENCE)
- span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_MODEL,
- request_model)
- span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_ENCODING_FORMATS,
- [kwargs.get('encoding_format', 'float')])
- span.set_attribute(SemanticConvetion.GEN_AI_RESPONSE_MODEL,
- request_model)
- span.set_attribute(SemanticConvetion.SERVER_ADDRESS,
- server_address)
- span.set_attribute(SemanticConvetion.SERVER_PORT,
- server_port)
- span.set_attribute(SemanticConvetion.GEN_AI_USAGE_INPUT_TOKENS,
- input_tokens)
-
- # Set Span attributes (Extras)
- span.set_attribute(DEPLOYMENT_ENVIRONMENT,
- environment)
- span.set_attribute(SERVICE_NAME,
- application_name)
- span.set_attribute(SemanticConvetion.GEN_AI_USAGE_TOTAL_TOKENS,
- input_tokens)
- span.set_attribute(SemanticConvetion.GEN_AI_USAGE_COST,
- cost)
- span.set_attribute(SemanticConvetion.GEN_AI_SDK_VERSION,
- version)
-
- if capture_message_content:
- span.add_event(
- name=SemanticConvetion.GEN_AI_CONTENT_PROMPT_EVENT,
- attributes={
- SemanticConvetion.GEN_AI_CONTENT_PROMPT: str(kwargs.get("input", "")),
- },
- )
-
- span.set_status(Status(StatusCode.OK))
-
- if disable_metrics is False:
- attributes = create_metrics_attributes(
- service_name=application_name,
- deployment_environment=environment,
- operation=SemanticConvetion.GEN_AI_OPERATION_TYPE_EMBEDDING,
- system=SemanticConvetion.GEN_AI_SYSTEM_AZURE_AI_INFERENCE,
- request_model=request_model,
- server_address=server_address,
- server_port=server_port,
- response_model=request_model,
- )
- metrics["genai_client_usage_tokens"].record(
- input_tokens, attributes
- )
- metrics["genai_client_operation_duration"].record(
- end_time - start_time, attributes
- )
- metrics["genai_requests"].add(1, attributes)
- metrics["genai_prompt_tokens"].add(input_tokens, attributes)
- metrics["genai_cost"].record(cost, attributes)
-
- # Return original response
- return response
-
- except Exception as e:
- handle_exception(span, e)
- logger.error("Error in trace creation: %s", e)
-
- # Return original response
- return response
+ response = process_chat_response(
+ response=response,
+ request_model=request_model,
+ pricing_info=pricing_info,
+ server_port=server_port,
+ server_address=server_address,
+ environment=environment,
+ application_name=application_name,
+ metrics=metrics,
+ event_provider=event_provider,
+ start_time=start_time,
+ span=span,
+ capture_message_content=capture_message_content,
+ disable_metrics=disable_metrics,
+ version=version,
+ **kwargs
+ )
+
+ return response
return wrapper
diff --git a/sdk/python/src/openlit/instrumentation/azure_ai_inference/azure_ai_inference.py b/sdk/python/src/openlit/instrumentation/azure_ai_inference/azure_ai_inference.py
index 14578044..81470d7b 100644
--- a/sdk/python/src/openlit/instrumentation/azure_ai_inference/azure_ai_inference.py
+++ b/sdk/python/src/openlit/instrumentation/azure_ai_inference/azure_ai_inference.py
@@ -4,18 +4,15 @@ Module for monitoring Azure AI Inference API calls.
import logging
import time
-from opentelemetry.trace import SpanKind, Status, StatusCode
-from opentelemetry.sdk.resources import SERVICE_NAME, TELEMETRY_SDK_NAME, DEPLOYMENT_ENVIRONMENT
+from opentelemetry.trace import SpanKind
from openlit.__helpers import (
- get_chat_model_cost,
- get_embed_model_cost,
handle_exception,
- response_as_dict,
- calculate_ttft,
- calculate_tbt,
- create_metrics_attributes,
set_server_address_and_port,
- general_tokens
+)
+from openlit.instrumentation.azure_ai_inference.utils import (
+ process_chunk,
+ process_chat_response,
+ process_streaming_chat_response,
)
from openlit.semcov import SemanticConvetion
@@ -23,37 +20,21 @@ from openlit.semcov import SemanticConvetion
logger = logging.getLogger(__name__)
def complete(version, environment, application_name,
- tracer, pricing_info, capture_message_content, metrics, disable_metrics):
+ tracer, event_provider, pricing_info, capture_message_content, metrics, disable_metrics):
"""
- Generates a telemetry wrapper for chat to collect metrics.
-
- Args:
- version: Version of the monitoring package.
- environment: Deployment environment (e.g., production, staging).
- application_name: Name of the application using the Azure AI Inference API.
- tracer: OpenTelemetry tracer for creating spans.
- pricing_info: Information used for calculating the cost of Azure AI Inference usage.
- capture_message_content: Flag indicating whether to trace the actual content.
-
- Returns:
- A function that wraps the chat method to add telemetry.
+ Generates a telemetry wrapper for GenAI function call
"""
class TracedSyncStream:
"""
- Wrapper for streaming responses to collect metrics and trace data.
- Wraps the response to collect message IDs and aggregated response.
-
- This class implements the '__aiter__' and '__anext__' methods that
- handle asynchronous streaming responses.
-
- This class also implements '__aenter__' and '__aexit__' methods that
- handle asynchronous context management protocol.
+ Wrapper for streaming responses to collect telemetry.
"""
+
def __init__(
self,
wrapped,
span,
+ span_name,
kwargs,
server_address,
server_port,
@@ -61,12 +42,13 @@ def complete(version, environment, application_name,
):
self.__wrapped__ = wrapped
self._span = span
- # Placeholder for aggregating streaming response
+ self._span_name = span_name
self._llmresponse = ""
self._response_id = ""
self._response_model = ""
self._finish_reason = ""
- self._system_fingerprint = ""
+ self._input_tokens = 0
+ self._output_tokens = 0
self._args = args
self._kwargs = kwargs
@@ -95,197 +77,33 @@ def complete(version, environment, application_name,
def __next__(self):
try:
chunk = self.__wrapped__.__next__()
- end_time = time.time()
- # Record the timestamp for the current chunk
- self._timestamps.append(end_time)
-
- if len(self._timestamps) == 1:
- # Calculate time to first chunk
- self._ttft = calculate_ttft(self._timestamps, self._start_time)
-
- chunked = response_as_dict(chunk)
- # Collect message IDs and aggregated response from events
- if (len(chunked.get('choices')) > 0 and ('delta' in chunked.get('choices')[0] and
- 'content' in chunked.get('choices')[0].get('delta'))):
-
- content = chunked.get('choices')[0].get('delta').get('content')
- if content:
- self._llmresponse += content
- self._response_id = chunked.get('id')
- self._response_model = chunked.get('model')
- self._finish_reason = chunked.get('choices')[0].get('finish_reason')
- self._system_fingerprint = chunked.get('system_fingerprint')
+ process_chunk(self, chunk)
return chunk
except StopIteration:
- # Handling exception ensure observability without disrupting operation
try:
- self._end_time = time.time()
- if len(self._timestamps) > 1:
- self._tbt = calculate_tbt(self._timestamps)
-
- # Format 'messages' into a single string
- message_prompt = self._kwargs.get("messages", "")
- formatted_messages = []
- for message in message_prompt:
- role = message["role"]
- content = message["content"]
-
- if isinstance(content, list):
- content_str_list = []
- for item in content:
- if item["type"] == "text":
- content_str_list.append(f'text: {item["text"]}')
- elif (item["type"] == "image_url" and
- not item["image_url"]["url"].startswith("data:")):
- content_str_list.append(f'image_url: {item["image_url"]["url"]}')
- content_str = ", ".join(content_str_list)
- formatted_messages.append(f"{role}: {content_str}")
- else:
- formatted_messages.append(f"{role}: {content}")
- prompt = "\n".join(formatted_messages)
-
- request_model = self._kwargs.get("model", "gpt-4o")
-
- # Calculate tokens using input prompt and aggregated response
- input_tokens = general_tokens(prompt)
- output_tokens = general_tokens(self._llmresponse)
-
- # Calculate cost of the operation
- cost = get_chat_model_cost(request_model,
- pricing_info, input_tokens,
- output_tokens)
-
- # Set Span attributes (OTel Semconv)
- self._span.set_attribute(TELEMETRY_SDK_NAME, "openlit")
- self._span.set_attribute(SemanticConvetion.GEN_AI_OPERATION,
- SemanticConvetion.GEN_AI_OPERATION_TYPE_CHAT)
- self._span.set_attribute(SemanticConvetion.GEN_AI_SYSTEM,
- SemanticConvetion.GEN_AI_SYSTEM_AZURE_AI_INFERENCE)
- self._span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_MODEL,
- request_model)
- self._span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_SEED,
- self._kwargs.get("seed", ""))
- self._span.set_attribute(SemanticConvetion.SERVER_PORT,
- self._server_port)
- self._span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_FREQUENCY_PENALTY,
- self._kwargs.get("frequency_penalty", 0.0))
- self._span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_MAX_TOKENS,
- self._kwargs.get("max_tokens", -1))
- self._span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_PRESENCE_PENALTY,
- self._kwargs.get("presence_penalty", 0.0))
- self._span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_STOP_SEQUENCES,
- self._kwargs.get("stop", []))
- self._span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_TEMPERATURE,
- self._kwargs.get("temperature", 1.0))
- self._span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_TOP_P,
- self._kwargs.get("top_p", 1.0))
- self._span.set_attribute(SemanticConvetion.GEN_AI_RESPONSE_FINISH_REASON,
- [self._finish_reason])
- self._span.set_attribute(SemanticConvetion.GEN_AI_RESPONSE_ID,
- self._response_id)
- self._span.set_attribute(SemanticConvetion.GEN_AI_RESPONSE_MODEL,
- self._response_model)
- self._span.set_attribute(SemanticConvetion.GEN_AI_USAGE_INPUT_TOKENS,
- input_tokens)
- self._span.set_attribute(SemanticConvetion.GEN_AI_USAGE_OUTPUT_TOKENS,
- output_tokens)
- self._span.set_attribute(SemanticConvetion.SERVER_ADDRESS,
- self._server_address)
- self._span.set_attribute(SemanticConvetion.GEN_AI_RESPONSE_SYSTEM_FINGERPRINT,
- self._system_fingerprint)
- if isinstance(self._llmresponse, str):
- self._span.set_attribute(SemanticConvetion.GEN_AI_OUTPUT_TYPE,
- "text")
- else:
- self._span.set_attribute(SemanticConvetion.GEN_AI_OUTPUT_TYPE,
- "json")
-
- # Set Span attributes (Extra)
- self._span.set_attribute(DEPLOYMENT_ENVIRONMENT,
- environment)
- self._span.set_attribute(SERVICE_NAME,
- application_name)
- self._span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_IS_STREAM,
- True)
- self._span.set_attribute(SemanticConvetion.GEN_AI_USAGE_TOTAL_TOKENS,
- input_tokens + output_tokens)
- self._span.set_attribute(SemanticConvetion.GEN_AI_USAGE_COST,
- cost)
- self._span.set_attribute(SemanticConvetion.GEN_AI_SERVER_TBT,
- self._tbt)
- self._span.set_attribute(SemanticConvetion.GEN_AI_SERVER_TTFT,
- self._ttft)
- self._span.set_attribute(SemanticConvetion.GEN_AI_SDK_VERSION,
- version)
- if capture_message_content:
- self._span.add_event(
- name=SemanticConvetion.GEN_AI_CONTENT_PROMPT_EVENT,
- attributes={
- SemanticConvetion.GEN_AI_CONTENT_PROMPT: prompt,
- },
- )
- self._span.add_event(
- name=SemanticConvetion.GEN_AI_CONTENT_COMPLETION_EVENT,
- attributes={
- SemanticConvetion.GEN_AI_CONTENT_COMPLETION: self._llmresponse,
- },
+ with tracer.start_as_current_span(self._span_name, kind= SpanKind.CLIENT) as self._span:
+ process_streaming_chat_response(
+ self,
+ pricing_info=pricing_info,
+ environment=environment,
+ application_name=application_name,
+ metrics=metrics,
+ event_provider=event_provider,
+ capture_message_content=capture_message_content,
+ disable_metrics=disable_metrics,
+ version=version
)
- self._span.set_status(Status(StatusCode.OK))
-
- if disable_metrics is False:
- attributes = create_metrics_attributes(
- service_name=application_name,
- deployment_environment=environment,
- operation=SemanticConvetion.GEN_AI_OPERATION_TYPE_CHAT,
- system=SemanticConvetion.GEN_AI_SYSTEM_AZURE_AI_INFERENCE,
- request_model=request_model,
- server_address=self._server_address,
- server_port=self._server_port,
- response_model=self._response_model,
- )
-
- metrics["genai_client_usage_tokens"].record(
- input_tokens + output_tokens, attributes
- )
- metrics["genai_client_operation_duration"].record(
- self._end_time - self._start_time, attributes
- )
- metrics["genai_server_tbt"].record(
- self._tbt, attributes
- )
- metrics["genai_server_ttft"].record(
- self._ttft, attributes
- )
- metrics["genai_requests"].add(1, attributes)
- metrics["genai_completion_tokens"].add(output_tokens, attributes)
- metrics["genai_prompt_tokens"].add(input_tokens, attributes)
- metrics["genai_cost"].record(cost, attributes)
except Exception as e:
handle_exception(self._span, e)
logger.error("Error in trace creation: %s", e)
- finally:
- self._span.end()
raise
def wrapper(wrapped, instance, args, kwargs):
"""
- Wraps the 'chat.completions' API call to add telemetry.
-
- This collects metrics such as execution time, cost, and token usage, and handles errors
- gracefully, adding details to the trace for observability.
-
- Args:
- wrapped: The original 'chat.completions' method to be wrapped.
- instance: The instance of the class where the original method is defined.
- args: Positional arguments for the 'chat.completions' method.
- kwargs: Keyword arguments for the 'chat.completions' method.
-
- Returns:
- The response from the original 'chat.completions' method.
+ Wraps the GenAI function call.
"""
- # Check if streaming is enabled for the API call
streaming = kwargs.get("stream", False)
server_address, server_port = set_server_address_and_port(instance, "models.github.ai", 443)
request_model = kwargs.get("model", "gpt-4o")
@@ -294,292 +112,33 @@ def complete(version, environment, application_name,
# pylint: disable=no-else-return
if streaming:
- # Special handling for streaming response to accommodate the nature of data flow
awaited_wrapped = wrapped(*args, **kwargs)
span = tracer.start_span(span_name, kind=SpanKind.CLIENT)
- return TracedSyncStream(awaited_wrapped, span, kwargs, server_address, server_port)
+ return TracedSyncStream(awaited_wrapped, span, span_name, kwargs, server_address, server_port)
- # Handling for non-streaming responses
else:
- with tracer.start_as_current_span(span_name, kind= SpanKind.CLIENT) as span:
+ with tracer.start_as_current_span(span_name, kind=SpanKind.CLIENT) as span:
start_time = time.time()
response = wrapped(*args, **kwargs)
- end_time = time.time()
-
- response_dict = response_as_dict(response)
-
- try:
- # Format 'messages' into a single string
- message_prompt = kwargs.get("messages", "")
- formatted_messages = []
- for message in message_prompt:
- role = message["role"]
- content = message["content"]
-
- if isinstance(content, list):
- content_str = ", ".join(
- f'{item["type"]}: {item["text"] if "text" in item else item["image_url"]}'
- if "type" in item else f'text: {item["text"]}'
- for item in content
- )
- formatted_messages.append(f"{role}: {content_str}")
- else:
- formatted_messages.append(f"{role}: {content}")
- prompt = "\n".join(formatted_messages)
-
- input_tokens = response_dict.get('usage').get('prompt_tokens')
- output_tokens = response_dict.get('usage').get('completion_tokens')
-
- # Calculate cost of the operation
- cost = get_chat_model_cost(request_model,
- pricing_info, input_tokens,
- output_tokens)
-
- # Set base span attribues (OTel Semconv)
- span.set_attribute(TELEMETRY_SDK_NAME, "openlit")
- span.set_attribute(SemanticConvetion.GEN_AI_OPERATION,
- SemanticConvetion.GEN_AI_OPERATION_TYPE_CHAT)
- span.set_attribute(SemanticConvetion.GEN_AI_SYSTEM,
- SemanticConvetion.GEN_AI_SYSTEM_AZURE_AI_INFERENCE)
- span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_MODEL,
- request_model)
- span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_SEED,
- kwargs.get("seed", ""))
- span.set_attribute(SemanticConvetion.SERVER_PORT,
- server_port)
- span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_FREQUENCY_PENALTY,
- kwargs.get("frequency_penalty", 0.0))
- span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_MAX_TOKENS,
- kwargs.get("max_tokens", -1))
- span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_PRESENCE_PENALTY,
- kwargs.get("presence_penalty", 0.0))
- span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_STOP_SEQUENCES,
- kwargs.get("stop", []))
- span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_TEMPERATURE,
- kwargs.get("temperature", 1.0))
- span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_TOP_P,
- kwargs.get("top_p", 1.0))
- span.set_attribute(SemanticConvetion.GEN_AI_RESPONSE_ID,
- response_dict.get("id"))
- span.set_attribute(SemanticConvetion.GEN_AI_RESPONSE_MODEL,
- response_dict.get('model'))
- span.set_attribute(SemanticConvetion.GEN_AI_USAGE_INPUT_TOKENS,
- input_tokens)
- span.set_attribute(SemanticConvetion.GEN_AI_USAGE_OUTPUT_TOKENS,
- output_tokens)
- span.set_attribute(SemanticConvetion.SERVER_ADDRESS,
- server_address)
- span.set_attribute(SemanticConvetion.GEN_AI_RESPONSE_SYSTEM_FINGERPRINT,
- response_dict.get('system_fingerprint'))
-
- # Set base span attribues (Extras)
- span.set_attribute(DEPLOYMENT_ENVIRONMENT,
- environment)
- span.set_attribute(SERVICE_NAME,
- application_name)
- span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_IS_STREAM,
- False)
- span.set_attribute(SemanticConvetion.GEN_AI_USAGE_TOTAL_TOKENS,
- input_tokens + output_tokens)
- span.set_attribute(SemanticConvetion.GEN_AI_USAGE_COST,
- cost)
- span.set_attribute(SemanticConvetion.GEN_AI_SERVER_TTFT,
- end_time - start_time)
- span.set_attribute(SemanticConvetion.GEN_AI_SDK_VERSION,
- version)
- if capture_message_content:
- span.add_event(
- name=SemanticConvetion.GEN_AI_CONTENT_PROMPT_EVENT,
- attributes={
- SemanticConvetion.GEN_AI_CONTENT_PROMPT: prompt,
- },
- )
-
- for i in range(kwargs.get('n',1)):
- span.set_attribute(SemanticConvetion.GEN_AI_RESPONSE_FINISH_REASON,
- [response_dict.get('choices')[i].get('finish_reason')])
- if capture_message_content:
- span.add_event(
- name=SemanticConvetion.GEN_AI_CONTENT_COMPLETION_EVENT,
- attributes={
- # pylint: disable=line-too-long
- SemanticConvetion.GEN_AI_CONTENT_COMPLETION: str(response_dict.get('choices')[i].get('message').get('content')),
- },
- )
- if kwargs.get('tools'):
- span.set_attribute(SemanticConvetion.GEN_AI_TOOL_CALLS,
- str(response_dict.get('choices')[i].get('message').get('tool_calls')))
-
- if isinstance(response_dict.get('choices')[i].get('message').get('content'), str):
- span.set_attribute(SemanticConvetion.GEN_AI_OUTPUT_TYPE,
- "text")
- elif response_dict.get('choices')[i].get('message').get('content') is not None:
- span.set_attribute(SemanticConvetion.GEN_AI_OUTPUT_TYPE,
- "json")
-
- span.set_status(Status(StatusCode.OK))
-
- if disable_metrics is False:
- attributes = create_metrics_attributes(
- service_name=application_name,
- deployment_environment=environment,
- operation=SemanticConvetion.GEN_AI_OPERATION_TYPE_CHAT,
- system=SemanticConvetion.GEN_AI_SYSTEM_AZURE_AI_INFERENCE,
- request_model=request_model,
- server_address=server_address,
- server_port=server_port,
- response_model=response_dict.get('model'),
- )
-
- metrics["genai_client_usage_tokens"].record(
- input_tokens + output_tokens, attributes
- )
- metrics["genai_client_operation_duration"].record(
- end_time - start_time, attributes
- )
- metrics["genai_server_ttft"].record(
- end_time - start_time, attributes
- )
- metrics["genai_requests"].add(1, attributes)
- metrics["genai_completion_tokens"].add(output_tokens, attributes)
- metrics["genai_prompt_tokens"].add(input_tokens, attributes)
- metrics["genai_cost"].record(cost, attributes)
-
- # Return original response
- return response
-
- except Exception as e:
- handle_exception(span, e)
- logger.error("Error in trace creation: %s", e)
-
- # Return original response
- return response
-
- return wrapper
-
-def embedding(version, environment, application_name,
- tracer, pricing_info, capture_message_content, metrics, disable_metrics):
- """
- Generates a telemetry wrapper for embeddings to collect metrics.
-
- Args:
- version: Version of the monitoring package.
- environment: Deployment environment (e.g., production, staging).
- application_name: Name of the application using the Azure Inference API.
- tracer: OpenTelemetry tracer for creating spans.
- pricing_info: Information used for calculating the cost of Azure Inference usage.
- capture_message_content: Flag indicating whether to trace the actual content.
-
- Returns:
- A function that wraps the embeddings method to add telemetry.
- """
-
- def wrapper(wrapped, instance, args, kwargs):
- """
- Wraps the 'embeddings' API call to add telemetry.
-
- This collects metrics such as execution time, cost, and token usage, and handles errors
- gracefully, adding details to the trace for observability.
-
- Args:
- wrapped: The original 'embeddings' method to be wrapped.
- instance: The instance of the class where the original method is defined.
- args: Positional arguments for the 'embeddings' method.
- kwargs: Keyword arguments for the 'embeddings' method.
-
- Returns:
- The response from the original 'embeddings' method.
- """
-
- server_address, server_port = set_server_address_and_port(instance, "models.github.ai", 443)
- request_model = kwargs.get("model", "text-embedding-ada-002")
-
- span_name = f"{SemanticConvetion.GEN_AI_OPERATION_TYPE_EMBEDDING} {request_model}"
-
- with tracer.start_as_current_span(span_name, kind= SpanKind.CLIENT) as span:
- start_time = time.time()
- response = wrapped(*args, **kwargs)
- end_time = time.time()
-
- response_dict = response_as_dict(response)
- try:
- input_tokens = response_dict.get('usage').get('prompt_tokens')
-
- # Calculate cost of the operation
- cost = get_embed_model_cost(request_model,
- pricing_info, input_tokens)
-
- # Set Span attributes (OTel Semconv)
- span.set_attribute(TELEMETRY_SDK_NAME, "openlit")
- span.set_attribute(SemanticConvetion.GEN_AI_OPERATION,
- SemanticConvetion.GEN_AI_OPERATION_TYPE_EMBEDDING)
- span.set_attribute(SemanticConvetion.GEN_AI_SYSTEM,
- SemanticConvetion.GEN_AI_SYSTEM_AZURE_AI_INFERENCE)
- span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_MODEL,
- request_model)
- span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_ENCODING_FORMATS,
- [kwargs.get('encoding_format', 'float')])
- span.set_attribute(SemanticConvetion.GEN_AI_RESPONSE_MODEL,
- request_model)
- span.set_attribute(SemanticConvetion.SERVER_ADDRESS,
- server_address)
- span.set_attribute(SemanticConvetion.SERVER_PORT,
- server_port)
- span.set_attribute(SemanticConvetion.GEN_AI_USAGE_INPUT_TOKENS,
- input_tokens)
-
- # Set Span attributes (Extras)
- span.set_attribute(DEPLOYMENT_ENVIRONMENT,
- environment)
- span.set_attribute(SERVICE_NAME,
- application_name)
- span.set_attribute(SemanticConvetion.GEN_AI_USAGE_TOTAL_TOKENS,
- input_tokens)
- span.set_attribute(SemanticConvetion.GEN_AI_USAGE_COST,
- cost)
- span.set_attribute(SemanticConvetion.GEN_AI_SDK_VERSION,
- version)
-
- if capture_message_content:
- span.add_event(
- name=SemanticConvetion.GEN_AI_CONTENT_PROMPT_EVENT,
- attributes={
- SemanticConvetion.GEN_AI_CONTENT_PROMPT: str(kwargs.get("input", "")),
- },
- )
-
- span.set_status(Status(StatusCode.OK))
-
- if disable_metrics is False:
- attributes = create_metrics_attributes(
- service_name=application_name,
- deployment_environment=environment,
- operation=SemanticConvetion.GEN_AI_OPERATION_TYPE_EMBEDDING,
- system=SemanticConvetion.GEN_AI_SYSTEM_AZURE_AI_INFERENCE,
- request_model=request_model,
- server_address=server_address,
- server_port=server_port,
- response_model=request_model,
- )
- metrics["genai_client_usage_tokens"].record(
- input_tokens, attributes
- )
- metrics["genai_client_operation_duration"].record(
- end_time - start_time, attributes
- )
- metrics["genai_requests"].add(1, attributes)
- metrics["genai_prompt_tokens"].add(input_tokens, attributes)
- metrics["genai_cost"].record(cost, attributes)
-
- # Return original response
- return response
-
- except Exception as e:
- handle_exception(span, e)
- logger.error("Error in trace creation: %s", e)
-
- # Return original response
- return response
+ response = process_chat_response(
+ response=response,
+ request_model=request_model,
+ pricing_info=pricing_info,
+ server_port=server_port,
+ server_address=server_address,
+ environment=environment,
+ application_name=application_name,
+ metrics=metrics,
+ event_provider=event_provider,
+ start_time=start_time,
+ span=span,
+ capture_message_content=capture_message_content,
+ disable_metrics=disable_metrics,
+ version=version,
+ **kwargs
+ )
+
+ return response
return wrapper
diff --git a/sdk/python/src/openlit/instrumentation/azure_ai_inference/utils.py b/sdk/python/src/openlit/instrumentation/azure_ai_inference/utils.py
new file mode 100644
index 00000000..5fc07078
--- /dev/null
+++ b/sdk/python/src/openlit/instrumentation/azure_ai_inference/utils.py
@@ -0,0 +1,225 @@
+"""
+Azure AI Inference OpenTelemetry instrumentation utility functions
+"""
+import time
+
+from opentelemetry.sdk.resources import SERVICE_NAME, TELEMETRY_SDK_NAME, DEPLOYMENT_ENVIRONMENT
+from opentelemetry.trace import Status, StatusCode
+
+from openlit.__helpers import (
+ calculate_ttft,
+ response_as_dict,
+ calculate_tbt,
+ extract_and_format_input,
+ get_chat_model_cost,
+ create_metrics_attributes,
+ otel_event,
+ concatenate_all_contents
+)
+from openlit.semcov import SemanticConvetion
+
+def process_chunk(self, chunk):
+ """
+ Process a chunk of response data and update state.
+ """
+
+ end_time = time.time()
+ # Record the timestamp for the current chunk
+ self._timestamps.append(end_time)
+
+ if len(self._timestamps) == 1:
+ # Calculate time to first chunk
+ self._ttft = calculate_ttft(self._timestamps, self._start_time)
+
+ chunked = response_as_dict(chunk)
+
+ # Collect message IDs and aggregated response from events
+ if (len(chunked.get('choices')) > 0 and ('delta' in chunked.get('choices')[0] and
+ 'content' in chunked.get('choices')[0].get('delta'))):
+
+ if content := chunked.get('choices')[0].get('delta').get('content'):
+ self._llmresponse += content
+
+ if chunked.get('choices')[0].get('finish_reason') is not None:
+ self._finish_reason = chunked.get('choices')[0].get('finish_reason')
+
+ if chunked.get('usage') is not None:
+ self._input_tokens = chunked.get('usage').get('prompt_tokens')
+ self._response_id = chunked.get('id')
+ self._response_model = chunked.get('model')
+ self._output_tokens = chunked.get('usage').get('completion_tokens')
+
+def common_chat_logic(scope, pricing_info, environment, application_name, metrics,
+ event_provider, capture_message_content, disable_metrics, version, is_stream):
+ """
+ Process chat request and generate Telemetry
+ """
+
+ scope._end_time = time.time()
+ if len(scope._timestamps) > 1:
+ scope._tbt = calculate_tbt(scope._timestamps)
+
+ formatted_messages = extract_and_format_input(scope._kwargs.get('messages', ''))
+ request_model = scope._kwargs.get('model', 'claude-3-opus-20240229')
+
+ cost = get_chat_model_cost(request_model, pricing_info, scope._input_tokens, scope._output_tokens)
+
+ # Set Span attributes (OTel Semconv)
+ scope._span.set_attribute(TELEMETRY_SDK_NAME, 'openlit')
+ scope._span.set_attribute(SemanticConvetion.GEN_AI_OPERATION, SemanticConvetion.GEN_AI_OPERATION_TYPE_CHAT)
+ scope._span.set_attribute(SemanticConvetion.GEN_AI_SYSTEM, SemanticConvetion.GEN_AI_SYSTEM_AZURE_AI_INFERENCE)
+ scope._span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_MODEL, request_model)
+ scope._span.set_attribute(SemanticConvetion.SERVER_PORT, scope._server_port)
+ scope._span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_MAX_TOKENS, scope._kwargs.get('max_tokens', -1))
+ scope._span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_STOP_SEQUENCES, scope._kwargs.get('stop', []))
+ scope._span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_TEMPERATURE, scope._kwargs.get('temperature', 1.0))
+ scope._span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_TOP_K, scope._kwargs.get('top_k', 1.0))
+ scope._span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_TOP_P, scope._kwargs.get('top_p', 1.0))
+ scope._span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_FREQUENCY_PENALTY,
+ scope._kwargs.get("frequency_penalty", 0.0))
+ scope._span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_PRESENCE_PENALTY,
+ scope._kwargs.get("presence_penalty", 0.0))
+ scope._span.set_attribute(SemanticConvetion.GEN_AI_RESPONSE_FINISH_REASON, [scope._finish_reason])
+ scope._span.set_attribute(SemanticConvetion.GEN_AI_RESPONSE_ID, scope._response_id)
+ scope._span.set_attribute(SemanticConvetion.GEN_AI_RESPONSE_MODEL, scope._response_model)
+ scope._span.set_attribute(SemanticConvetion.GEN_AI_USAGE_INPUT_TOKENS, scope._input_tokens)
+ scope._span.set_attribute(SemanticConvetion.GEN_AI_USAGE_OUTPUT_TOKENS, scope._output_tokens)
+ scope._span.set_attribute(SemanticConvetion.SERVER_ADDRESS, scope._server_address)
+
+ scope._span.set_attribute(SemanticConvetion.GEN_AI_OUTPUT_TYPE,
+ 'text' if isinstance(scope._llmresponse, str) else 'json')
+
+ scope._span.set_attribute(DEPLOYMENT_ENVIRONMENT, environment)
+ scope._span.set_attribute(SERVICE_NAME, application_name)
+ scope._span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_IS_STREAM, is_stream)
+ scope._span.set_attribute(SemanticConvetion.GEN_AI_CLIENT_TOKEN_USAGE, scope._input_tokens + scope._output_tokens)
+ scope._span.set_attribute(SemanticConvetion.GEN_AI_USAGE_COST, cost)
+ scope._span.set_attribute(SemanticConvetion.GEN_AI_SERVER_TBT, scope._tbt)
+ scope._span.set_attribute(SemanticConvetion.GEN_AI_SERVER_TTFT, scope._ttft)
+ scope._span.set_attribute(SemanticConvetion.GEN_AI_SDK_VERSION, version)
+
+ # To be removed one the change to log events (from span events) is complete
+ prompt = concatenate_all_contents(formatted_messages)
+ if capture_message_content:
+ scope._span.add_event(
+ name=SemanticConvetion.GEN_AI_CONTENT_PROMPT_EVENT,
+ attributes={
+ SemanticConvetion.GEN_AI_CONTENT_PROMPT: prompt,
+ },
+ )
+ scope._span.add_event(
+ name=SemanticConvetion.GEN_AI_CONTENT_COMPLETION_EVENT,
+ attributes={
+ SemanticConvetion.GEN_AI_CONTENT_COMPLETION: scope._llmresponse,
+ },
+ )
+
+ choice_event_body = {
+ 'finish_reason': scope._finish_reason,
+ 'index': 0,
+ 'message': {
+ **({'content': scope._llmresponse} if capture_message_content else {}),
+ 'role': 'assistant'
+ }
+ }
+
+ # Emit events
+ for role in ['user', 'system', 'assistant', 'tool']:
+ if formatted_messages.get(role, {}).get('content', ''):
+ event = otel_event(
+ name=getattr(SemanticConvetion, f'GEN_AI_{role.upper()}_MESSAGE'),
+ attributes={
+ SemanticConvetion.GEN_AI_SYSTEM: SemanticConvetion.GEN_AI_SYSTEM_AZURE_AI_INFERENCE
+ },
+ body = {
+ # pylint: disable=line-too-long
+ **({'content': formatted_messages.get(role, {}).get('content', '')} if capture_message_content else {}),
+ 'role': formatted_messages.get(role, {}).get('role', []),
+ **({
+ 'tool_calls': {
+ 'function': {
+ # pylint: disable=line-too-long
+ 'name': (scope._tool_calls[0].get('function', {}).get('name', '') if scope._tool_calls else ''),
+ 'arguments': (scope._tool_calls[0].get('function', {}).get('arguments', '') if scope._tool_calls else '')
+ },
+ 'id': (scope._tool_calls[0].get('id', '') if scope._tool_calls else ''),
+ 'type': 'function'
+ }
+ } if role == 'assistant' else {}),
+ **({
+ 'id': (scope._tool_calls[0].get('id', '') if scope._tool_calls else '')
+ } if role == 'tool' else {})
+ }
+ )
+ event_provider.emit(event)
+
+ choice_event = otel_event(
+ name=SemanticConvetion.GEN_AI_CHOICE,
+ attributes={
+ SemanticConvetion.GEN_AI_SYSTEM: SemanticConvetion.GEN_AI_SYSTEM_AZURE_AI_INFERENCE
+ },
+ body=choice_event_body
+ )
+ event_provider.emit(choice_event)
+
+ scope._span.set_status(Status(StatusCode.OK))
+
+ if not disable_metrics:
+ metrics_attributes = create_metrics_attributes(
+ service_name=application_name,
+ deployment_environment=environment,
+ operation=SemanticConvetion.GEN_AI_OPERATION_TYPE_CHAT,
+ system=SemanticConvetion.GEN_AI_SYSTEM_AZURE_AI_INFERENCE,
+ request_model=request_model,
+ server_address=scope._server_address,
+ server_port=scope._server_port,
+ response_model=scope._response_model,
+ )
+
+ metrics['genai_client_usage_tokens'].record(scope._input_tokens + scope._output_tokens, metrics_attributes)
+ metrics['genai_client_operation_duration'].record(scope._end_time - scope._start_time, metrics_attributes)
+ metrics['genai_server_tbt'].record(scope._tbt, metrics_attributes)
+ metrics['genai_server_ttft'].record(scope._ttft, metrics_attributes)
+ metrics['genai_requests'].add(1, metrics_attributes)
+ metrics['genai_completion_tokens'].add(scope._output_tokens, metrics_attributes)
+ metrics['genai_prompt_tokens'].add(scope._input_tokens, metrics_attributes)
+ metrics['genai_cost'].record(cost, metrics_attributes)
+
+def process_streaming_chat_response(self, pricing_info, environment, application_name, metrics,
+ event_provider, capture_message_content=False, disable_metrics=False, version=''):
+ """
+ Process chat request and generate Telemetry
+ """
+
+ common_chat_logic(self, pricing_info, environment, application_name, metrics,
+ event_provider, capture_message_content, disable_metrics, version, is_stream=True)
+
+def process_chat_response(response, request_model, pricing_info, server_port, server_address,
+ environment, application_name, metrics, event_provider, start_time,
+ span, capture_message_content=False, disable_metrics=False, version='1.0.0', **kwargs):
+ """
+ Process chat request and generate Telemetry
+ """
+
+ self = type('GenericScope', (), {})()
+ response_dict = response_as_dict(response)
+
+ # pylint: disable = no-member
+ self._start_time = start_time
+ self._end_time = time.time()
+ self._span = span
+ self._llmresponse = response_dict.get('choices', {})[0].get('message', '').get('content', '')
+ self._input_tokens = response_dict.get('usage').get('prompt_tokens')
+ self._output_tokens = response_dict.get('usage').get('completion_tokens')
+ self._response_model = response_dict.get('model', '')
+ self._finish_reason = response_dict.get('choices', {})[0].get('finish_reason', '')
+ self._response_id = response_dict.get('id', '')
+ self._timestamps = []
+ self._ttft, self._tbt = self._end_time - self._start_time, 0
+ self._server_address, self._server_port = server_address, server_port
+ self._kwargs = kwargs
+
+ common_chat_logic(self, pricing_info, environment, application_name, metrics,
+ event_provider, capture_message_content, disable_metrics, version, is_stream=False)
+
+ return response