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]: Update Cohere Instrumentation to align with OTel Semconv (#618)

Patcher1ca4cd32fb3e
diff --git a/sdk/python/pyproject.toml b/sdk/python/pyproject.toml
index 9e20da8b..dde0383f 100644
--- a/sdk/python/pyproject.toml
+++ b/sdk/python/pyproject.toml
@@ -1,6 +1,6 @@
[tool.poetry]
name = "openlit"
-version = "1.33.8"
+version = "1.33.9"
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/anthropic/anthropic.py b/sdk/python/src/openlit/instrumentation/anthropic/anthropic.py
index 7698fb06..f4bb61ce 100644
--- a/sdk/python/src/openlit/instrumentation/anthropic/anthropic.py
+++ b/sdk/python/src/openlit/instrumentation/anthropic/anthropic.py
@@ -148,7 +148,7 @@ def messages(version, environment, application_name, tracer,
formatted_messages.append(f"{role}: {content}")
prompt = "\n".join(formatted_messages)
- request_model = self._kwargs.get("model", "gpt-4o")
+ request_model = self._kwargs.get("model", "claude-3-5-sonnet-latest")
# Calculate cost of the operation
cost = get_chat_model_cost(request_model,
@@ -312,7 +312,6 @@ def messages(version, environment, application_name, tracer,
if isinstance(content, list):
content_str = ", ".join(
- # pylint: disable=line-too-long
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
@@ -397,7 +396,6 @@ def messages(version, environment, application_name, tracer,
span.add_event(
name=SemanticConvetion.GEN_AI_CONTENT_COMPLETION_EVENT,
attributes={
- # pylint: disable=line-too-long
SemanticConvetion.GEN_AI_CONTENT_COMPLETION: llm_response,
},
)
diff --git a/sdk/python/src/openlit/instrumentation/anthropic/async_anthropic.py b/sdk/python/src/openlit/instrumentation/anthropic/async_anthropic.py
index 6cab60fe..7969193b 100644
--- a/sdk/python/src/openlit/instrumentation/anthropic/async_anthropic.py
+++ b/sdk/python/src/openlit/instrumentation/anthropic/async_anthropic.py
@@ -148,7 +148,7 @@ def async_messages(version, environment, application_name, tracer,
formatted_messages.append(f"{role}: {content}")
prompt = "\n".join(formatted_messages)
- request_model = self._kwargs.get("model", "gpt-4o")
+ request_model = self._kwargs.get("model", "claude-3-5-sonnet-latest")
# Calculate cost of the operation
cost = get_chat_model_cost(request_model,
@@ -312,7 +312,6 @@ def async_messages(version, environment, application_name, tracer,
if isinstance(content, list):
content_str = ", ".join(
- # pylint: disable=line-too-long
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
@@ -397,7 +396,6 @@ def async_messages(version, environment, application_name, tracer,
span.add_event(
name=SemanticConvetion.GEN_AI_CONTENT_COMPLETION_EVENT,
attributes={
- # pylint: disable=line-too-long
SemanticConvetion.GEN_AI_CONTENT_COMPLETION: llm_response,
},
)
diff --git a/sdk/python/src/openlit/instrumentation/cohere/__init__.py b/sdk/python/src/openlit/instrumentation/cohere/__init__.py
index 65ae8da6..89f728de 100644
--- a/sdk/python/src/openlit/instrumentation/cohere/__init__.py
+++ b/sdk/python/src/openlit/instrumentation/cohere/__init__.py
@@ -6,8 +6,10 @@ from opentelemetry.instrumentation.instrumentor import BaseInstrumentor
from wrapt import wrap_function_wrapper
from openlit.instrumentation.cohere.cohere import chat, chat_stream, embed
+from openlit.instrumentation.cohere.async_cohere import async_chat, async_chat_stream, async_embed
-_instruments = ("cohere >= 5.3.2",)
+
+_instruments = ("cohere >= 5.14.0",)
class CohereInstrumentor(BaseInstrumentor):
"""An instrumentor for Cohere's client library."""
@@ -25,24 +27,45 @@ class CohereInstrumentor(BaseInstrumentor):
disable_metrics = kwargs.get("disable_metrics")
version = importlib.metadata.version("cohere")
+ # Sync Client
+ wrap_function_wrapper(
+ "cohere.client_v2",
+ "ClientV2.chat",
+ chat(version, environment, application_name,
+ tracer, pricing_info, trace_content, metrics, disable_metrics),
+ )
+ wrap_function_wrapper(
+ "cohere.client_v2",
+ "ClientV2.chat_stream",
+ chat_stream(version, environment, application_name,
+ tracer, pricing_info, trace_content, metrics, disable_metrics),
+ )
+ wrap_function_wrapper(
+ "cohere.client_v2",
+ "ClientV2.embed",
+ embed(version, environment, application_name,
+ tracer, pricing_info, trace_content, metrics, disable_metrics),
+ )
+
+ # Async Client
wrap_function_wrapper(
- "cohere.client",
- "Client.chat",
- chat("cohere.chat", version, environment, application_name,
+ "cohere.client_v2",
+ "AsyncClientV2.chat",
+ async_chat(version, environment, application_name,
tracer, pricing_info, trace_content, metrics, disable_metrics),
)
wrap_function_wrapper(
- "cohere.client",
- "Client.chat_stream",
- chat_stream("cohere.chat", version, environment, application_name,
+ "cohere.client_v2",
+ "AsyncClientV2.chat_stream",
+ async_chat_stream(version, environment, application_name,
tracer, pricing_info, trace_content, metrics, disable_metrics),
)
wrap_function_wrapper(
- "cohere.client",
- "Client.embed",
- embed("cohere.embed", version, environment, application_name,
+ "cohere.client_v2",
+ "AsyncClientV2.embed",
+ async_embed(version, environment, application_name,
tracer, pricing_info, trace_content, metrics, disable_metrics),
)
diff --git a/sdk/python/src/openlit/instrumentation/cohere/async_cohere.py b/sdk/python/src/openlit/instrumentation/cohere/async_cohere.py
new file mode 100644
index 00000000..def04ed2
--- /dev/null
+++ b/sdk/python/src/openlit/instrumentation/cohere/async_cohere.py
@@ -0,0 +1,611 @@
+"""
+Module for monitoring Cohere 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 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
+)
+from openlit.semcov import SemanticConvetion
+
+# Initialize logger for logging potential issues and operations
+logger = logging.getLogger(__name__)
+
+def async_embed(version, environment, application_name, tracer,
+ pricing_info, trace_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 Cohere API.
+ tracer: OpenTelemetry tracer for creating spans.
+ pricing_info: Information used for calculating the cost of Cohere usage.
+ trace_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 'embed' 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 'embed' method to be wrapped.
+ instance: The instance of the class where the original method is defined.
+ args: Positional arguments for the 'embed' method.
+ kwargs: Keyword arguments for the 'embed' method.
+
+ Returns:
+ The response from the original 'embed' method.
+ """
+
+ server_address, server_port = set_server_address_and_port(instance, "api.cohere.com", 443)
+ request_model = kwargs.get("model", "mbed-english-v3.0")
+
+ 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('meta').get('billed_units').get('input_tokens')
+ # Calculate cost of the operation
+ cost = get_embed_model_cost(kwargs.get("model", "embed-english-v2.0"),
+ 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_COHERE)
+ span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_MODEL,
+ request_model)
+ span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_ENCODING_FORMATS,
+ kwargs.get('embedding_types', ['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)
+ span.set_attribute(SemanticConvetion.GEN_AI_OUTPUT_TYPE,
+ response_dict.get('response_type'))
+
+ # 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 trace_content:
+ span.add_event(
+ name=SemanticConvetion.GEN_AI_CONTENT_PROMPT_EVENT,
+ attributes={
+ SemanticConvetion.GEN_AI_CONTENT_PROMPT: str(kwargs.get("texts", "")),
+ },
+ )
+
+ 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_COHERE,
+ 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
+
+ return wrapper
+
+def async_chat(version, environment, application_name, tracer,
+ pricing_info, trace_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 Cohere API.
+ tracer: OpenTelemetry tracer for creating spans.
+ pricing_info: Information used for calculating the cost of Cohere usage.
+ trace_content: Flag indicating whether to trace the actual content.
+
+ Returns:
+ A function that wraps the chat method to add telemetry.
+ """
+
+ async def wrapper(wrapped, instance, args, kwargs):
+ """
+ Wraps the 'chat' 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' method to be wrapped.
+ instance: The instance of the class where the original method is defined.
+ args: Positional arguments for the 'chat' method.
+ kwargs: Keyword arguments for the 'chat' method.
+
+ Returns:
+ The response from the original 'chat' method.
+ """
+
+ server_address, server_port = set_server_address_and_port(instance, "api.cohere.com", 443)
+ request_model = kwargs.get("model", "command-r-plus-08-2024")
+
+ span_name = f"{SemanticConvetion.GEN_AI_OPERATION_TYPE_CHAT} {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:
+ # 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('billed_units').get('input_tokens')
+ output_tokens = response_dict.get('usage').get('billed_units').get('output_tokens')
+
+ # Calculate cost of the operation
+ cost = get_chat_model_cost(request_model, pricing_info,
+ input_tokens, output_tokens)
+
+ llm_response = response_dict.get('message').get('content')[0].get('text')
+
+ # 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_COHERE)
+ 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_sequences", []))
+ span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_TEMPERATURE,
+ kwargs.get("temperature", 0.3))
+ span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_TOP_P,
+ kwargs.get("k", 1.0))
+ span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_TOP_P,
+ kwargs.get("p", 1.0))
+ span.set_attribute(SemanticConvetion.GEN_AI_RESPONSE_ID,
+ response_dict.get("id"))
+ span.set_attribute(SemanticConvetion.GEN_AI_RESPONSE_MODEL,
+ request_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)
+ if isinstance(llm_response, str):
+ span.set_attribute(SemanticConvetion.GEN_AI_OUTPUT_TYPE,
+ "text")
+ else:
+ span.set_attribute(SemanticConvetion.GEN_AI_OUTPUT_TYPE,
+ "json")
+
+ # 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 trace_content:
+ span.add_event(
+ name=SemanticConvetion.GEN_AI_CONTENT_PROMPT_EVENT,
+ attributes={
+ SemanticConvetion.GEN_AI_CONTENT_PROMPT: prompt,
+ },
+ )
+ span.add_event(
+ name=SemanticConvetion.GEN_AI_CONTENT_COMPLETION_EVENT,
+ attributes={
+ SemanticConvetion.GEN_AI_CONTENT_COMPLETION: llm_response,
+ },
+ )
+
+ 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_COHERE,
+ request_model=request_model,
+ server_address=server_address,
+ server_port=server_port,
+ response_model=request_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_chat_stream(version, environment, application_name,
+ tracer, pricing_info, trace_content, metrics, disable_metrics):
+ """
+ Generates a telemetry wrapper for chat_stream to collect metrics.
+
+ Args:
+ gen_ai_endpoint: Endpoint identifier for logging and tracing.
+ version: Version of the monitoring package.
+ environment: Deployment environment (e.g., production, staging).
+ application_name: Name of the application using the Cohere API.
+ tracer: OpenTelemetry tracer for creating spans.
+ pricing_info: Information used for calculating the cost of Cohere usage.
+ trace_content: Flag indicating whether to trace the actual content.
+
+ Returns:
+ A function that wraps the chat method to add telemetry.
+ """
+
+ async def wrapper(wrapped, instance, args, kwargs):
+ """
+ Wraps the 'chat_stream' 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_stream' method to be wrapped.
+ instance: The instance of the class where the original method is defined.
+ args: Positional arguments for the 'chat_stream' method.
+ kwargs: Keyword arguments for the 'chat_stream' method.
+
+ Returns:
+ The response from the original 'chat_stream' method.
+ """
+
+ class TracedAsyncStream:
+ """
+ Wrapper for streaming responses to collect metrics and trace data.
+ Wraps the 'cohere.AsyncStream' 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.
+ """
+ def __init__(
+ self,
+ wrapped,
+ span,
+ kwargs,
+ server_address,
+ server_port,
+ **args,
+ ):
+ self.__wrapped__ = wrapped
+ self._span = span
+ # Placeholder for aggregating streaming response
+ self._llmresponse = ""
+ self._response_id = ""
+ self._finish_reason = ""
+ self._input_tokens = ""
+ self._output_tokens = ""
+
+ self._args = args
+ self._kwargs = kwargs
+ self._start_time = time.time()
+ self._end_time = None
+ self._timestamps = []
+ self._ttft = 0
+ self._tbt = 0
+ self._server_address = server_address
+ self._server_port = server_port
+
+ async def __aenter__(self):
+ await self.__wrapped__.__aenter__()
+ return self
+
+ async def __aexit__(self, exc_type, exc_value, traceback):
+ await self.__wrapped__.__aexit__(exc_type, exc_value, traceback)
+
+ def __aiter__(self):
+ return self
+
+ async def __getattr__(self, name):
+ """Delegate attribute access to the wrapped object."""
+ return getattr(await self.__wrapped__, 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)
+
+ if chunked.get('type') == 'message-start':
+ self._response_id = chunked.get('id')
+
+ if chunked.get('type') == 'content-delta':
+ content = chunked.get('delta').get('message').get('text')
+ if content:
+ self._llmresponse += content
+
+ if chunked.get('type') == 'message-end':
+ self._finish_reason = chunked.get('delta').get('finish_reason')
+ self._input_tokens = chunked.get('delta').get('usage').get('billed_units').get('input_tokens')
+ self._output_tokens = chunked.get('delta').get('usage').get('billed_units').get('output_tokens')
+
+ 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", "command-r-plus")
+
+ # Calculate cost of the operation
+ cost = get_chat_model_cost(request_model,
+ pricing_info, self._input_tokens,
+ self._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_COHERE)
+ 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_sequences", []))
+ self._span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_TEMPERATURE,
+ self._kwargs.get("temperature", 0.3))
+ self._span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_TOP_K,
+ self._kwargs.get("k", 1.0))
+ self._span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_TOP_P,
+ self._kwargs.get("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,
+ request_model)
+ self._span.set_attribute(SemanticConvetion.GEN_AI_USAGE_INPUT_TOKENS,
+ self._input_tokens)
+ self._span.set_attribute(SemanticConvetion.GEN_AI_USAGE_OUTPUT_TOKENS,
+ self._output_tokens)
+ self._span.set_attribute(SemanticConvetion.SERVER_ADDRESS,
+ self._server_address)
+
+ 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,
+ self._input_tokens + self._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 trace_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,
+ },
+ )
+ 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_COHERE,
+ request_model=request_model,
+ server_address=self._server_address,
+ server_port=self._server_port,
+ response_model=request_model,
+ )
+
+ metrics["genai_client_usage_tokens"].record(
+ self._input_tokens + self._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(self._output_tokens, attributes)
+ metrics["genai_prompt_tokens"].add(self._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
+
+ server_address, server_port = set_server_address_and_port(instance, "api.cohere.com", 443)
+ request_model = kwargs.get("model", "command-r-plus")
+
+ span_name = f"{SemanticConvetion.GEN_AI_OPERATION_TYPE_CHAT} {request_model}"
+
+ 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 wrapper
diff --git a/sdk/python/src/openlit/instrumentation/cohere/cohere.py b/sdk/python/src/openlit/instrumentation/cohere/cohere.py
index b78bd9db..9b772f61 100644
--- a/sdk/python/src/openlit/instrumentation/cohere/cohere.py
+++ b/sdk/python/src/openlit/instrumentation/cohere/cohere.py
@@ -1,29 +1,37 @@
-# pylint: disable=duplicate-code, broad-exception-caught, too-many-statements, unused-argument, possibly-used-before-assignment
"""
Module for monitoring Cohere API calls.
"""
import logging
+import time
from opentelemetry.trace import SpanKind, Status, StatusCode
-from opentelemetry.sdk.resources import TELEMETRY_SDK_NAME
-from openlit.__helpers import get_chat_model_cost, get_embed_model_cost, handle_exception
+from opentelemetry.sdk.resources import SERVICE_NAME, TELEMETRY_SDK_NAME, DEPLOYMENT_ENVIRONMENT
+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
+)
from openlit.semcov import SemanticConvetion
# Initialize logger for logging potential issues and operations
logger = logging.getLogger(__name__)
-def embed(gen_ai_endpoint, version, environment, application_name, tracer,
+def embed(version, environment, application_name, tracer,
pricing_info, trace_content, metrics, disable_metrics):
"""
Generates a telemetry wrapper for embeddings to collect metrics.
Args:
- gen_ai_endpoint: Endpoint identifier for logging and tracing.
version: Version of the monitoring package.
environment: Deployment environment (e.g., production, staging).
- application_name: Name of the application using the OpenAI API.
+ application_name: Name of the application using the Cohere API.
tracer: OpenTelemetry tracer for creating spans.
- pricing_info: Information used for calculating the cost of OpenAI usage.
+ pricing_info: Information used for calculating the cost of Cohere usage.
trace_content: Flag indicating whether to trace the actual content.
Returns:
@@ -47,80 +55,85 @@ def embed(gen_ai_endpoint, version, environment, application_name, tracer,
The response from the original 'embed' method.
"""
- with tracer.start_as_current_span(gen_ai_endpoint, kind= SpanKind.CLIENT) as span:
- response = wrapped(*args, **kwargs)
+ server_address, server_port = set_server_address_and_port(instance, "api.cohere.com", 443)
+ request_model = kwargs.get("model", "mbed-english-v3.0")
- try:
- # Get prompt from kwargs and store as a single string
- prompt = " ".join(kwargs.get("texts", []))
+ 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('meta').get('billed_units').get('input_tokens')
# Calculate cost of the operation
cost = get_embed_model_cost(kwargs.get("model", "embed-english-v2.0"),
- pricing_info,
- response.meta.billed_units.input_tokens)
+ pricing_info, input_tokens)
- # Set Span attributes
+ # Set Span attributes (OTel Semconv)
span.set_attribute(TELEMETRY_SDK_NAME, "openlit")
- span.set_attribute(SemanticConvetion.GEN_AI_SYSTEM,
- SemanticConvetion.GEN_AI_SYSTEM_COHERE)
span.set_attribute(SemanticConvetion.GEN_AI_OPERATION,
SemanticConvetion.GEN_AI_OPERATION_TYPE_EMBEDDING)
- span.set_attribute(SemanticConvetion.GEN_AI_ENDPOINT,
- gen_ai_endpoint)
- span.set_attribute(SemanticConvetion.GEN_AI_ENVIRONMENT,
- environment)
- span.set_attribute(SemanticConvetion.GEN_AI_APPLICATION_NAME,
- application_name)
+ span.set_attribute(SemanticConvetion.GEN_AI_SYSTEM,
+ SemanticConvetion.GEN_AI_SYSTEM_COHERE)
span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_MODEL,
- kwargs.get("model", "embed-english-v2.0"))
+ request_model)
span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_ENCODING_FORMATS,
- kwargs.get("embedding_types", "float"))
- span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_EMBEDDING_DIMENSION,
- kwargs.get("input_type", ""))
- span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_USER,
- kwargs.get("user", ""))
- span.set_attribute(SemanticConvetion.GEN_AI_RESPONSE_ID,
- response.id)
+ kwargs.get('embedding_types', ['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,
- response.meta.billed_units.input_tokens)
+ input_tokens)
+ span.set_attribute(SemanticConvetion.GEN_AI_OUTPUT_TYPE,
+ response_dict.get('response_type'))
+
+ # 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,
- response.meta.billed_units.input_tokens)
+ input_tokens)
span.set_attribute(SemanticConvetion.GEN_AI_USAGE_COST,
cost)
+ span.set_attribute(SemanticConvetion.GEN_AI_SDK_VERSION,
+ version)
+
if trace_content:
span.add_event(
name=SemanticConvetion.GEN_AI_CONTENT_PROMPT_EVENT,
attributes={
- SemanticConvetion.GEN_AI_CONTENT_PROMPT: prompt,
+ SemanticConvetion.GEN_AI_CONTENT_PROMPT: str(kwargs.get("texts", "")),
},
)
span.set_status(Status(StatusCode.OK))
if disable_metrics is False:
- attributes = {
- TELEMETRY_SDK_NAME:
- "openlit",
- SemanticConvetion.GEN_AI_APPLICATION_NAME:
- application_name,
- SemanticConvetion.GEN_AI_SYSTEM:
- SemanticConvetion.GEN_AI_SYSTEM_COHERE,
- SemanticConvetion.GEN_AI_ENVIRONMENT:
- environment,
- SemanticConvetion.GEN_AI_OPERATION:
- SemanticConvetion.GEN_AI_OPERATION_TYPE_EMBEDDING,
- SemanticConvetion.GEN_AI_REQUEST_MODEL:
- kwargs.get("model", "embed-english-v2.0")
- }
-
- metrics["genai_requests"].add(1, attributes)
- metrics["genai_total_tokens"].add(
- response.meta.billed_units.input_tokens, attributes
+ attributes = create_metrics_attributes(
+ service_name=application_name,
+ deployment_environment=environment,
+ operation=SemanticConvetion.GEN_AI_OPERATION_TYPE_EMBEDDING,
+ system=SemanticConvetion.GEN_AI_SYSTEM_COHERE,
+ request_model=request_model,
+ server_address=server_address,
+ server_port=server_port,
+ response_model=request_model,
)
- metrics["genai_prompt_tokens"].add(
- response.meta.billed_units.input_tokens, attributes
+ 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
@@ -135,18 +148,17 @@ def embed(gen_ai_endpoint, version, environment, application_name, tracer,
return wrapper
-def chat(gen_ai_endpoint, version, environment, application_name, tracer,
+def chat(version, environment, application_name, tracer,
pricing_info, trace_content, metrics, disable_metrics):
"""
Generates a telemetry wrapper for chat to collect metrics.
Args:
- gen_ai_endpoint: Endpoint identifier for logging and tracing.
version: Version of the monitoring package.
environment: Deployment environment (e.g., production, staging).
- application_name: Name of the application using the OpenAI API.
+ application_name: Name of the application using the Cohere API.
tracer: OpenTelemetry tracer for creating spans.
- pricing_info: Information used for calculating the cost of OpenAI usage.
+ pricing_info: Information used for calculating the cost of Cohere usage.
trace_content: Flag indicating whether to trace the actual content.
Returns:
@@ -170,96 +182,145 @@ def chat(gen_ai_endpoint, version, environment, application_name, tracer,
The response from the original 'chat' method.
"""
- with tracer.start_as_current_span(gen_ai_endpoint, kind=SpanKind.CLIENT) as span:
+ server_address, server_port = set_server_address_and_port(instance, "api.cohere.com", 443)
+ request_model = kwargs.get("model", "command-r-plus-08-2024")
+
+ span_name = f"{SemanticConvetion.GEN_AI_OPERATION_TYPE_CHAT} {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:
+ # 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('billed_units').get('input_tokens')
+ output_tokens = response_dict.get('usage').get('billed_units').get('output_tokens')
+
# Calculate cost of the operation
- cost = get_chat_model_cost(kwargs.get("model", "command"),
- pricing_info,
- response.meta.billed_units.input_tokens,
- response.meta.billed_units.output_tokens)
+ cost = get_chat_model_cost(request_model, pricing_info,
+ input_tokens, output_tokens)
+
+ llm_response = response_dict.get('message').get('content')[0].get('text')
- # Set Span attributes
+ # Set base span attribues (OTel Semconv)
span.set_attribute(TELEMETRY_SDK_NAME, "openlit")
- span.set_attribute(SemanticConvetion.GEN_AI_SYSTEM,
- SemanticConvetion.GEN_AI_SYSTEM_COHERE)
span.set_attribute(SemanticConvetion.GEN_AI_OPERATION,
SemanticConvetion.GEN_AI_OPERATION_TYPE_CHAT)
- span.set_attribute(SemanticConvetion.GEN_AI_ENDPOINT,
- gen_ai_endpoint)
- span.set_attribute(SemanticConvetion.GEN_AI_ENVIRONMENT,
- environment)
- span.set_attribute(SemanticConvetion.GEN_AI_APPLICATION_NAME,
- application_name)
+ span.set_attribute(SemanticConvetion.GEN_AI_SYSTEM,
+ SemanticConvetion.GEN_AI_SYSTEM_COHERE)
span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_MODEL,
- kwargs.get("model", "command"))
- span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_TEMPERATURE,
- kwargs.get("temperature", 0.3))
- span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_MAX_TOKENS,
- kwargs.get("max_tokens", -1))
+ 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_IS_STREAM,
- False)
+ span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_STOP_SEQUENCES,
+ kwargs.get("stop_sequences", []))
+ span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_TEMPERATURE,
+ kwargs.get("temperature", 0.3))
+ span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_TOP_P,
+ kwargs.get("k", 1.0))
+ span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_TOP_P,
+ kwargs.get("p", 1.0))
span.set_attribute(SemanticConvetion.GEN_AI_RESPONSE_ID,
- response.generation_id)
- span.set_attribute(SemanticConvetion.GEN_AI_RESPONSE_FINISH_REASON,
- [response.finish_reason])
+ response_dict.get("id"))
+ span.set_attribute(SemanticConvetion.GEN_AI_RESPONSE_MODEL,
+ request_model)
span.set_attribute(SemanticConvetion.GEN_AI_USAGE_INPUT_TOKENS,
- response.meta.billed_units.input_tokens)
+ input_tokens)
span.set_attribute(SemanticConvetion.GEN_AI_USAGE_OUTPUT_TOKENS,
- response.meta.billed_units.output_tokens)
+ output_tokens)
+ span.set_attribute(SemanticConvetion.SERVER_ADDRESS,
+ server_address)
+ if isinstance(llm_response, str):
+ span.set_attribute(SemanticConvetion.GEN_AI_OUTPUT_TYPE,
+ "text")
+ else:
+ span.set_attribute(SemanticConvetion.GEN_AI_OUTPUT_TYPE,
+ "json")
+
+ # 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,
- response.meta.billed_units.input_tokens +
- response.meta.billed_units.output_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 trace_content:
span.add_event(
name=SemanticConvetion.GEN_AI_CONTENT_PROMPT_EVENT,
attributes={
- SemanticConvetion.GEN_AI_CONTENT_PROMPT: kwargs.get("message", ""),
+ SemanticConvetion.GEN_AI_CONTENT_PROMPT: prompt,
},
)
span.add_event(
name=SemanticConvetion.GEN_AI_CONTENT_COMPLETION_EVENT,
attributes={
- SemanticConvetion.GEN_AI_CONTENT_COMPLETION: response.text,
+ SemanticConvetion.GEN_AI_CONTENT_COMPLETION: llm_response,
},
)
span.set_status(Status(StatusCode.OK))
if disable_metrics is False:
- attributes = {
- TELEMETRY_SDK_NAME:
- "openlit",
- SemanticConvetion.GEN_AI_APPLICATION_NAME:
- application_name,
- SemanticConvetion.GEN_AI_SYSTEM:
- SemanticConvetion.GEN_AI_SYSTEM_COHERE,
- SemanticConvetion.GEN_AI_ENVIRONMENT:
- environment,
- SemanticConvetion.GEN_AI_OPERATION:
- SemanticConvetion.GEN_AI_OPERATION_TYPE_CHAT,
- SemanticConvetion.GEN_AI_REQUEST_MODEL:
- kwargs.get("model", "command")
- }
+ attributes = create_metrics_attributes(
+ service_name=application_name,
+ deployment_environment=environment,
+ operation=SemanticConvetion.GEN_AI_OPERATION_TYPE_CHAT,
+ system=SemanticConvetion.GEN_AI_SYSTEM_COHERE,
+ request_model=request_model,
+ server_address=server_address,
+ server_port=server_port,
+ response_model=request_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_total_tokens"].add(
- response.meta.billed_units.input_tokens +
- response.meta.billed_units.output_tokens, attributes)
- metrics["genai_completion_tokens"].add(
- response.meta.billed_units.output_tokens, attributes)
- metrics["genai_prompt_tokens"].add(
- response.meta.billed_units.input_tokens, 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
@@ -274,7 +335,7 @@ def chat(gen_ai_endpoint, version, environment, application_name, tracer,
return wrapper
-def chat_stream(gen_ai_endpoint, version, environment, application_name,
+def chat_stream(version, environment, application_name,
tracer, pricing_info, trace_content, metrics, disable_metrics):
"""
Generates a telemetry wrapper for chat_stream to collect metrics.
@@ -283,9 +344,9 @@ def chat_stream(gen_ai_endpoint, version, environment, application_name,
gen_ai_endpoint: Endpoint identifier for logging and tracing.
version: Version of the monitoring package.
environment: Deployment environment (e.g., production, staging).
- application_name: Name of the application using the OpenAI API.
+ application_name: Name of the application using the Cohere API.
tracer: OpenTelemetry tracer for creating spans.
- pricing_info: Information used for calculating the cost of OpenAI usage.
+ pricing_info: Information used for calculating the cost of Cohere usage.
trace_content: Flag indicating whether to trace the actual content.
Returns:
@@ -309,111 +370,242 @@ def chat_stream(gen_ai_endpoint, version, environment, application_name,
The response from the original 'chat_stream' method.
"""
- def stream_generator():
- with tracer.start_as_current_span(gen_ai_endpoint, kind= SpanKind.CLIENT) as span:
+ class TracedSyncStream:
+ """
+ Wrapper for streaming responses to collect metrics and trace data.
+ Wraps the 'cohere.AsyncStream' 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.
+ """
+ def __init__(
+ self,
+ wrapped,
+ span,
+ kwargs,
+ server_address,
+ server_port,
+ **args,
+ ):
+ self.__wrapped__ = wrapped
+ self._span = span
# Placeholder for aggregating streaming response
- llmresponse = ""
-
- # Loop through streaming events capturing relevant details
- for event in wrapped(*args, **kwargs):
- # Collect message IDs and aggregated response from events
- if event.event_type == "stream-end":
- llmresponse = event.response.text
- prompt_tokens = event.response.meta.billed_units.input_tokens
- completion_tokens = event.response.meta.billed_units.output_tokens
- finish_reason = event.finish_reason
- if event.event_type == "stream-start":
- response_id = event.generation_id
- yield event
-
- # Handling exception ensure observability without disrupting operation
+ self._llmresponse = ""
+ self._response_id = ""
+ self._finish_reason = ""
+ self._input_tokens = ""
+ self._output_tokens = ""
+
+ self._args = args
+ self._kwargs = kwargs
+ self._start_time = time.time()
+ self._end_time = None
+ self._timestamps = []
+ self._ttft = 0
+ self._tbt = 0
+ self._server_address = server_address
+ self._server_port = server_port
+
+ def __enter__(self):
+ self.__wrapped__.__enter__()
+ return self
+
+ def __exit__(self, exc_type, exc_value, traceback):
+ self.__wrapped__.__exit__(exc_type, exc_value, traceback)
+
+ def __iter__(self):
+ return self
+
+ def __getattr__(self, name):
+ """Delegate attribute access to the wrapped object."""
+ return getattr(self.__wrapped__, name)
+
+ def __next__(self):
try:
- # Calculate cost of the operation
- cost = get_chat_model_cost(kwargs.get("model", "command"),
- pricing_info, prompt_tokens, completion_tokens)
-
- # Set Span attributes
- span.set_attribute(TELEMETRY_SDK_NAME, "openlit")
- span.set_attribute(SemanticConvetion.GEN_AI_SYSTEM,
- SemanticConvetion.GEN_AI_SYSTEM_COHERE)
- span.set_attribute(SemanticConvetion.GEN_AI_OPERATION,
- SemanticConvetion.GEN_AI_OPERATION_TYPE_CHAT)
- span.set_attribute(SemanticConvetion.GEN_AI_ENDPOINT,
- gen_ai_endpoint)
- span.set_attribute(SemanticConvetion.GEN_AI_ENVIRONMENT,
- environment)
- span.set_attribute(SemanticConvetion.GEN_AI_APPLICATION_NAME,
- application_name)
- span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_MODEL,
- kwargs.get("model", "command"))
- span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_TEMPERATURE,
- kwargs.get("temperature", 0.3))
- span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_MAX_TOKENS,
- kwargs.get("max_tokens", -1))
- span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_SEED,
- kwargs.get("seed", ""))
- span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_FREQUENCY_PENALTY,
- kwargs.get("frequency_penalty", 0.0))
- span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_PRESENCE_PENALTY,
- kwargs.get("presence_penalty", 0.0))
- span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_IS_STREAM,
- True)
- span.set_attribute(SemanticConvetion.GEN_AI_RESPONSE_ID,
- response_id)
- span.set_attribute(SemanticConvetion.GEN_AI_RESPONSE_FINISH_REASON,
- [finish_reason])
- span.set_attribute(SemanticConvetion.GEN_AI_USAGE_INPUT_TOKENS,
- prompt_tokens)
- span.set_attribute(SemanticConvetion.GEN_AI_USAGE_OUTPUT_TOKENS,
- completion_tokens)
- span.set_attribute(SemanticConvetion.GEN_AI_USAGE_TOTAL_TOKENS,
- prompt_tokens + completion_tokens)
- span.set_attribute(SemanticConvetion.GEN_AI_USAGE_COST,
- cost)
- if trace_content:
- span.add_event(
- name=SemanticConvetion.GEN_AI_CONTENT_PROMPT_EVENT,
- attributes={
- SemanticConvetion.GEN_AI_CONTENT_PROMPT: kwargs.get("message", ""),
- },
- )
- span.add_event(
- name=SemanticConvetion.GEN_AI_CONTENT_COMPLETION_EVENT,
- attributes={
- SemanticConvetion.GEN_AI_CONTENT_COMPLETION: llmresponse,
- },
- )
-
- span.set_status(Status(StatusCode.OK))
-
- if disable_metrics is False:
- attributes = {
- TELEMETRY_SDK_NAME:
- "openlit",
- SemanticConvetion.GEN_AI_APPLICATION_NAME:
- application_name,
- SemanticConvetion.GEN_AI_SYSTEM:
- SemanticConvetion.GEN_AI_SYSTEM_COHERE,
- SemanticConvetion.GEN_AI_ENVIRONMENT:
- environment,
- SemanticConvetion.GEN_AI_OPERATION:
- SemanticConvetion.GEN_AI_OPERATION_TYPE_CHAT,
- SemanticConvetion.GEN_AI_REQUEST_MODEL:
- kwargs.get("model", "command")
- }
-
- metrics["genai_requests"].add(1, attributes)
- metrics["genai_total_tokens"].add(
- prompt_tokens + completion_tokens, attributes
- )
- metrics["genai_completion_tokens"].add(completion_tokens, attributes)
- metrics["genai_prompt_tokens"].add(prompt_tokens, attributes)
- metrics["genai_cost"].record(cost, attributes)
-
- except Exception as e:
- handle_exception(span, e)
- logger.error("Error in trace creation: %s", e)
-
- return stream_generator()
+ 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)
+
+ if chunked.get('type') == 'message-start':
+ self._response_id = chunked.get('id')
+
+ if chunked.get('type') == 'content-delta':
+ content = chunked.get('delta').get('message').get('text')
+ if content:
+ self._llmresponse += content
+
+ if chunked.get('type') == 'message-end':
+ self._finish_reason = chunked.get('delta').get('finish_reason')
+ self._input_tokens = chunked.get('delta').get('usage').get('billed_units').get('input_tokens')
+ self._output_tokens = chunked.get('delta').get('usage').get('billed_units').get('output_tokens')
+
+ 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", "command-r-plus")
+
+ # Calculate cost of the operation
+ cost = get_chat_model_cost(request_model,
+ pricing_info, self._input_tokens,
+ self._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_COHERE)
+ 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_sequences", []))
+ self._span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_TEMPERATURE,
+ self._kwargs.get("temperature", 0.3))
+ self._span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_TOP_K,
+ self._kwargs.get("k", 1.0))
+ self._span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_TOP_P,
+ self._kwargs.get("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,
+ request_model)
+ self._span.set_attribute(SemanticConvetion.GEN_AI_USAGE_INPUT_TOKENS,
+ self._input_tokens)
+ self._span.set_attribute(SemanticConvetion.GEN_AI_USAGE_OUTPUT_TOKENS,
+ self._output_tokens)
+ self._span.set_attribute(SemanticConvetion.SERVER_ADDRESS,
+ self._server_address)
+
+ 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,
+ self._input_tokens + self._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 trace_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,
+ },
+ )
+ 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_COHERE,
+ request_model=request_model,
+ server_address=self._server_address,
+ server_port=self._server_port,
+ response_model=request_model,
+ )
+
+ metrics["genai_client_usage_tokens"].record(
+ self._input_tokens + self._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(self._output_tokens, attributes)
+ metrics["genai_prompt_tokens"].add(self._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
+
+ server_address, server_port = set_server_address_and_port(instance, "api.cohere.com", 443)
+ request_model = kwargs.get("model", "command-r-plus")
+
+ span_name = f"{SemanticConvetion.GEN_AI_OPERATION_TYPE_CHAT} {request_model}"
+
+ 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 wrapper
diff --git a/sdk/python/src/openlit/instrumentation/openai/async_openai.py b/sdk/python/src/openlit/instrumentation/openai/async_openai.py
index b844a2fe..19739159 100644
--- a/sdk/python/src/openlit/instrumentation/openai/async_openai.py
+++ b/sdk/python/src/openlit/instrumentation/openai/async_openai.py
@@ -390,9 +390,9 @@ def async_chat_completions(version, environment, application_name,
response_dict.get('system_fingerprint'))
# Set base span attribues (Extras)
- span.set_attribute(SemanticConvetion.GEN_AI_ENVIRONMENT,
+ span.set_attribute(DEPLOYMENT_ENVIRONMENT,
environment)
- span.set_attribute(SemanticConvetion.GEN_AI_APPLICATION_NAME,
+ span.set_attribute(SERVICE_NAME,
application_name)
span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_USER,
kwargs.get("user", ""))
@@ -522,7 +522,6 @@ def async_embedding(version, environment, application_name,
response_dict = response_as_dict(response)
try:
- request_model = kwargs.get("model", "text-embedding-ada-002")
input_tokens = response_dict.get('usage').get('prompt_tokens')
# Calculate cost of the operation
@@ -538,9 +537,9 @@ def async_embedding(version, environment, application_name,
span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_MODEL,
request_model)
span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_ENCODING_FORMATS,
- kwargs.get('encoding_format', 'float'))
+ [kwargs.get('encoding_format', 'float')])
span.set_attribute(SemanticConvetion.GEN_AI_RESPONSE_MODEL,
- response_dict.get('model'))
+ request_model)
span.set_attribute(SemanticConvetion.SERVER_ADDRESS,
server_address)
span.set_attribute(SemanticConvetion.SERVER_PORT,
@@ -589,7 +588,6 @@ def async_embedding(version, environment, application_name,
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)
diff --git a/sdk/python/src/openlit/instrumentation/openai/openai.py b/sdk/python/src/openlit/instrumentation/openai/openai.py
index 2baaa1c0..1ec61b2f 100644
--- a/sdk/python/src/openlit/instrumentation/openai/openai.py
+++ b/sdk/python/src/openlit/instrumentation/openai/openai.py
@@ -390,9 +390,9 @@ def chat_completions(version, environment, application_name,
response_dict.get('system_fingerprint'))
# Set base span attribues (Extras)
- span.set_attribute(SemanticConvetion.GEN_AI_ENVIRONMENT,
+ span.set_attribute(DEPLOYMENT_ENVIRONMENT,
environment)
- span.set_attribute(SemanticConvetion.GEN_AI_APPLICATION_NAME,
+ span.set_attribute(SERVICE_NAME,
application_name)
span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_USER,
kwargs.get("user", ""))
@@ -522,7 +522,6 @@ def embedding(version, environment, application_name,
response_dict = response_as_dict(response)
try:
- request_model = kwargs.get("model", "text-embedding-ada-002")
input_tokens = response_dict.get('usage').get('prompt_tokens')
# Calculate cost of the operation
@@ -538,9 +537,9 @@ def embedding(version, environment, application_name,
span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_MODEL,
request_model)
span.set_attribute(SemanticConvetion.GEN_AI_REQUEST_ENCODING_FORMATS,
- kwargs.get('encoding_format', 'float'))
+ [kwargs.get('encoding_format', 'float')])
span.set_attribute(SemanticConvetion.GEN_AI_RESPONSE_MODEL,
- response_dict.get('model'))
+ request_model)
span.set_attribute(SemanticConvetion.SERVER_ADDRESS,
server_address)
span.set_attribute(SemanticConvetion.SERVER_PORT,
@@ -589,7 +588,6 @@ def embedding(version, environment, application_name,
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)