add content opt-in env var

Author: Liudmila Molkova

CONTRIBUTING.md

@@ -227,7 +227,7 @@ tox -e py312-test-instrumentation-aiopg
 Some of the tox targets install packages from the [OpenTelemetry Python Core Repository](https://github.com/open-telemetry/opentelemetry-python) via pip. The version of the packages installed defaults to the main branch in that repository when tox is run locally. It is possible to install packages tagged with a specific git commit hash by setting an environment variable before running tox as per the following example:
 
 ```sh
-CORE_REPO_SHA=966750aad4ba2a9c123b4f14785958ffe50b54ae tox
+CORE_REPO_SHA=c49ad57bfe35cfc69bfa863d74058ca9bec55fc3 tox
 ```
 
 The continuous integration overrides that environment variable with as per the configuration [here](https://github.com/open-telemetry/opentelemetry-python-contrib/blob/main/.github/workflows/test_0.yml#L14).

instrumentation/opentelemetry-instrumentation-openai-v2/CHANGELOG.md

@@ -0,0 +1,15 @@
+# Changelog
+
+All notable changes to this project will be documented in this file.
+
+The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
+and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
+
+## Unreleased
+
+- Update OpenAI instrumentation to Semantic Conventions v1.28.0: add new attributes
+  and switch prompts and completions to log-based events.
+  ([#2925](https://github.com/open-telemetry/opentelemetry-python-contrib/pull/2925))
+
+- Initial OpenAI instrumentation
+  ([#2759](https://github.com/open-telemetry/opentelemetry-python-contrib/pull/2759))

instrumentation/opentelemetry-instrumentation-openai-v2/README.rst

@@ -1,12 +1,12 @@
 OpenTelemetry OpenAI Instrumentation
-====================================    
+====================================
 
 |pypi|
 
 .. |pypi| image:: https://badge.fury.io/py/opentelemetry-instrumentation-openai-v2.svg
    :target: https://pypi.org/project/opentelemetry-instrumentation-openai-v2/
 
-Instrumentation with OpenAI that supports the openai library and is
+Instrumentation with OpenAI that supports the OpenAI library and is
 specified to trace_integration using 'OpenAI'.
 
 

instrumentation/opentelemetry-instrumentation-openai-v2/src/opentelemetry/instrumentation/openai_v2/__init__.py

@@ -47,6 +47,7 @@
 from opentelemetry._events import get_event_logger
 from opentelemetry.instrumentation.instrumentor import BaseInstrumentor
 from opentelemetry.instrumentation.openai_v2.package import _instruments
+from opentelemetry.instrumentation.openai_v2.utils import is_content_enabled
 from opentelemetry.instrumentation.utils import unwrap
 from opentelemetry.semconv.schemas import Schemas
 from opentelemetry.trace import get_tracer
@@ -74,10 +75,13 @@ def _instrument(self, **kwargs):
             schema_url=Schemas.V1_28_0.value,
             event_logger_provider=event_provider,
         )
+
         wrap_function_wrapper(
             module="openai.resources.chat.completions",
             name="Completions.create",
-            wrapper=chat_completions_create(tracer, event_logger),
+            wrapper=chat_completions_create(
+                tracer, event_logger, is_content_enabled()
+            ),
         )
 
     def _uninstrument(self, **kwargs):

instrumentation/opentelemetry-instrumentation-openai-v2/src/opentelemetry/instrumentation/openai_v2/patch.py

@@ -37,7 +37,9 @@
 )
 
 
-def chat_completions_create(tracer: Tracer, event_logger: EventLogger):
+def chat_completions_create(
+    tracer: Tracer, event_logger: EventLogger, capture_content: bool
+):
     """Wrap the `create` method of the `ChatCompletion` class to trace it."""
 
     def traced_method(wrapped, instance, args, kwargs):
@@ -53,20 +55,22 @@ def traced_method(wrapped, instance, args, kwargs):
             if span.is_recording():
                 for message in kwargs.get("messages", []):
                     event_logger.emit(
-                        message_to_event(message, span.get_span_context())
+                        message_to_event(
+                            message, span.get_span_context(), capture_content
+                        )
                     )
 
             try:
                 result = wrapped(*args, **kwargs)
                 if is_streaming(kwargs):
                     return StreamWrapper(
-                        result,
-                        span,
-                        event_logger,
+                        result, span, event_logger, capture_content
                     )
                 else:
                     if span.is_recording():
-                        _set_response_attributes(span, result, event_logger)
+                        _set_response_attributes(
+                            span, result, event_logger, capture_content
+                        )
                     span.end()
                     return result
 
@@ -83,7 +87,9 @@ def traced_method(wrapped, instance, args, kwargs):
 
 
 @silently_fail
-def _set_response_attributes(span, result, event_logger: EventLogger):
+def _set_response_attributes(
+    span, result, event_logger: EventLogger, capture_content: bool
+):
     if not span.is_recording():
         return
 
@@ -95,7 +101,9 @@ def _set_response_attributes(span, result, event_logger: EventLogger):
         choices = result.choices
         for choice in choices:
             event_logger.emit(
-                choice_to_event(choice, span_ctx=span.get_span_context())
+                choice_to_event(
+                    choice, span.get_span_context(), capture_content
+                )
             )
 
         finish_reasons = []
@@ -182,11 +190,13 @@ def __init__(
         stream: Stream,
         span: Span,
         event_logger: EventLogger,
+        capture_content: bool,
     ):
         self.stream = stream
         self.span = span
         self.choice_buffers = []
         self._span_started = False
+        self.capture_content = capture_content
 
         self.event_logger = event_logger
         self.setup()
@@ -238,18 +248,20 @@ def cleanup(self):
                 choice = self.choice_buffers[c]
 
                 message = {"role": "assistant"}
-                if choice.text_content:
+                if self.capture_content and choice.text_content:
                     message["content"] = "".join(choice.text_content)
                 if choice.tool_calls_buffers:
                     tool_calls = []
                     for tool_call in choice.tool_calls_buffers:
+                        function = {"name": tool_call.function_name}
+                        if self.capture_content:
+                            function["arguments"] = "".join(
+                                tool_call.arguments
+                            )
                         tool_call_dict = {
                             "id": tool_call.tool_call_id,
                             "type": "function",
-                            "function": {
-                                "name": tool_call.function_name,
-                                "arguments": "".join(tool_call.arguments),
-                            },
+                            "function": function,
                         }
                         tool_calls.append(tool_call_dict)
                     message["tool_calls"] = tool_calls

instrumentation/opentelemetry-instrumentation-openai-v2/src/opentelemetry/instrumentation/openai_v2/utils.py

@@ -13,7 +13,8 @@
 # limitations under the License.
 
 import logging
-from typing import Any, Optional, Tuple, Union
+from os import environ
+from typing import Optional, Union
 from urllib.parse import urlparse
 
 from httpx import URL
@@ -27,6 +28,18 @@
     server_attributes as ServerAttributes,
 )
 
+OTEL_INSTRUMENTATION_OPENAI_CAPTURE_MESSAGE_CONTENT = (
+    "OTEL_INSTRUMENTATION_OPENAI_CAPTURE_MESSAGE_CONTENT"
+)
+
+
+def is_content_enabled() -> bool:
+    capture_content = environ.get(
+        OTEL_INSTRUMENTATION_OPENAI_CAPTURE_MESSAGE_CONTENT, None
+    )
+
+    return bool(capture_content)
+
 
 def silently_fail(func):
     """
@@ -48,20 +61,7 @@ def wrapper(*args, **kwargs):
     return wrapper
 
 
-def extract_content_from_prompt_message(message) -> Tuple[str, Any]:
-    content = get_property_value(message, "content")
-    if content:
-        return ("content", content)
-
-    tool_calls = extract_tool_calls(message)
-    if tool_calls:
-        return ("tool_calls", tool_calls)
-
-    else:
-        return None
-
-
-def extract_tool_calls(item):
+def extract_tool_calls(item, capture_content):
     tool_calls = get_property_value(item, "tool_calls")
     if tool_calls is None:
         return
@@ -82,7 +82,7 @@ def extract_tool_calls(item):
                 tool_call_dict["function"]["name"] = name
 
             arguments = get_property_value(func, "arguments")
-            if arguments:
+            if capture_content and arguments:
                 tool_call_dict["function"]["arguments"] = arguments.replace(
                     "\n", ""
                 )
@@ -120,18 +120,18 @@ def get_property_value(obj, property_name):
     return getattr(obj, property_name, None)
 
 
-def message_to_event(message, span_ctx):
+def message_to_event(message, span_ctx, capture_content):
     attributes = {
         GenAIAttributes.GEN_AI_SYSTEM: GenAIAttributes.GenAiSystemValues.OPENAI.value
     }
     role = get_property_value(message, "role")
     content = get_property_value(message, "content")
 
     body = {}
-    if content:
+    if capture_content and content:
         body["content"] = content
     if role == "assistant":
-        tool_calls = extract_tool_calls(message)
+        tool_calls = extract_tool_calls(message, capture_content)
         if tool_calls:
             body = {"tool_calls": tool_calls}
     elif role == "tool":
@@ -142,14 +142,14 @@ def message_to_event(message, span_ctx):
     return Event(
         name=f"gen_ai.{role}.message",
         attributes=attributes,
-        body=body,
+        body=body if body else None,
         trace_id=span_ctx.trace_id,
         span_id=span_ctx.span_id,
         trace_flags=span_ctx.trace_flags,
     )
 
 
-def choice_to_event(choice, span_ctx):
+def choice_to_event(choice, span_ctx, capture_content):
     attributes = {
         GenAIAttributes.GEN_AI_SYSTEM: GenAIAttributes.GenAiSystemValues.OPENAI.value
     }
@@ -165,11 +165,11 @@ def choice_to_event(choice, span_ctx):
             if choice.message and choice.message.role
             else "assistant"
         }
-        tool_calls = extract_tool_calls(choice.message)
+        tool_calls = extract_tool_calls(choice.message, capture_content)
         if tool_calls:
             message["tool_calls"] = tool_calls
         content = get_property_value(choice.message, "content")
-        if content:
+        if capture_content and content:
             message["content"] = content
         body["message"] = message