fix(openai): dynamically import types for 1.99 (#3244)

Author: dinmukhamedm

packages/opentelemetry-instrumentation-openai/opentelemetry/instrumentation/openai/shared/chat_wrappers.py

@@ -7,6 +7,7 @@
 from typing import List, Optional, Union
 
 from opentelemetry import context as context_api
+import pydantic
 from opentelemetry.instrumentation.openai.shared import (
     OPENAI_LLM_USAGE_TOKEN_TYPES,
     _get_openai_base_url,
@@ -50,9 +51,6 @@
 from opentelemetry.trace.status import Status, StatusCode
 from wrapt import ObjectProxy
 
-from openai.types.chat import ChatCompletionMessageToolCall
-from openai.types.chat.chat_completion_message import FunctionCall
-
 SPAN_NAME = "openai.chat"
 PROMPT_FILTER_KEY = "prompt_filter_results"
 CONTENT_FILTER_KEY = "content_filter_results"
@@ -961,8 +959,10 @@ async def _abuild_from_streaming_response(
     span.end()
 
 
+# pydantic.BaseModel here is ChatCompletionMessageFunctionToolCall (as of openai 1.99.7)
+# but we keep to a parent type to support older versions
 def _parse_tool_calls(
-    tool_calls: Optional[List[Union[dict, ChatCompletionMessageToolCall]]],
+    tool_calls: Optional[List[Union[dict, pydantic.BaseModel]]],
 ) -> Union[List[ToolCall], None]:
     """
     Util to correctly parse the tool calls data from the OpenAI API to this module's
@@ -976,12 +976,11 @@ def _parse_tool_calls(
     for tool_call in tool_calls:
         tool_call_data = None
 
-        # Handle dict or ChatCompletionMessageToolCall
         if isinstance(tool_call, dict):
             tool_call_data = copy.deepcopy(tool_call)
-        elif isinstance(tool_call, ChatCompletionMessageToolCall):
+        elif _is_chat_message_function_tool_call(tool_call):
             tool_call_data = tool_call.model_dump()
-        elif isinstance(tool_call, FunctionCall):
+        elif _is_function_call(tool_call):
             function_call = tool_call.model_dump()
             tool_call_data = ToolCall(
                 id="",
@@ -996,6 +995,34 @@ def _parse_tool_calls(
     return result
 
 
+def _is_chat_message_function_tool_call(model: Union[dict, pydantic.BaseModel]) -> bool:
+    try:
+        from openai.types.chat.chat_completion_message_function_tool_call import (
+            ChatCompletionMessageFunctionToolCall,
+        )
+
+        return isinstance(model, ChatCompletionMessageFunctionToolCall)
+    except Exception:
+        try:
+            # Since OpenAI 1.99.3, ChatCompletionMessageToolCall is a Union,
+            # and the isinstance check will fail. This is fine, because in all
+            # those versions, the check above will succeed.
+            from openai.types.chat.chat_completion_message_tool_call import (
+                ChatCompletionMessageToolCall,
+            )
+            return isinstance(model, ChatCompletionMessageToolCall)
+        except Exception:
+            return False
+
+
+def _is_function_call(model: Union[dict, pydantic.BaseModel]) -> bool:
+    try:
+        from openai.types.chat.chat_completion_message import FunctionCall
+        return isinstance(model, FunctionCall)
+    except Exception:
+        return False
+
+
 @singledispatch
 def _parse_choice_event(choice) -> ChoiceEvent:
     has_message = choice.message is not None

packages/opentelemetry-instrumentation-openai/opentelemetry/instrumentation/openai/v1/responses_wrappers.py

@@ -447,6 +447,14 @@ def responses_get_or_create_wrapper(tracer: Tracer, wrapped, instance, args, kwa
     merged_tools = existing_data.get("tools", []) + request_tools
 
     try:
+        parsed_response_output_text = None
+        if hasattr(parsed_response, "output_text"):
+            parsed_response_output_text = parsed_response.output_text
+        else:
+            try:
+                parsed_response_output_text = parsed_response.output[0].content[0].text
+            except Exception:
+                pass
         traced_data = TracedData(
             start_time=existing_data.get("start_time", start_time),
             response_id=parsed_response.id,
@@ -456,7 +464,7 @@ def responses_get_or_create_wrapper(tracer: Tracer, wrapped, instance, args, kwa
             output_blocks={block.id: block for block in parsed_response.output}
             | existing_data.get("output_blocks", {}),
             usage=existing_data.get("usage", parsed_response.usage),
-            output_text=existing_data.get("output_text", parsed_response.output_text),
+            output_text=existing_data.get("output_text", parsed_response_output_text),
             request_model=existing_data.get("request_model", kwargs.get("model")),
             response_model=existing_data.get("response_model", parsed_response.model),
         )
@@ -541,6 +549,15 @@ async def async_responses_get_or_create_wrapper(
     merged_tools = existing_data.get("tools", []) + request_tools
 
     try:
+        parsed_response_output_text = None
+        if hasattr(parsed_response, "output_text"):
+            parsed_response_output_text = parsed_response.output_text
+        else:
+            try:
+                parsed_response_output_text = parsed_response.output[0].content[0].text
+            except Exception:
+                pass
+
         traced_data = TracedData(
             start_time=existing_data.get("start_time", start_time),
             response_id=parsed_response.id,
@@ -550,7 +567,7 @@ async def async_responses_get_or_create_wrapper(
             output_blocks={block.id: block for block in parsed_response.output}
             | existing_data.get("output_blocks", {}),
             usage=existing_data.get("usage", parsed_response.usage),
-            output_text=existing_data.get("output_text", parsed_response.output_text),
+            output_text=existing_data.get("output_text", parsed_response_output_text),
             request_model=existing_data.get("request_model", kwargs.get("model")),
             response_model=existing_data.get("response_model", parsed_response.model),
         )

packages/opentelemetry-instrumentation-openai/poetry.lock

@@ -38,7 +38,7 @@ pytest = "^8.2.2"
 pytest-sugar = "1.0.0"
 vcrpy = "^6.0.1"
 pytest-recording = "^0.13.1"
-openai = { extras = ["datalib"], version = ">=1.66.0" }
+openai = { extras = ["datalib"], version = "1.99.7" }
 opentelemetry-sdk = "^1.27.0"
 pytest-asyncio = "^0.23.7"
 requests = "^2.31.0"

packages/opentelemetry-instrumentation-openai/pyproject.toml

@@ -4,8 +4,7 @@
 import httpx
 import pytest
 from openai.types.chat.chat_completion_message_tool_call import (
-    ChatCompletionMessageToolCall,
-    Function,
+    ChatCompletionMessageFunctionToolCall,
 )
 from opentelemetry.sdk._logs import LogData
 from opentelemetry.semconv._incubating.attributes import (
@@ -375,13 +374,21 @@ def test_chat_tool_calls_with_events_with_no_content(
 def test_chat_pydantic_based_tool_calls(
     instrument_legacy, span_exporter, log_exporter, openai_client
 ):
+    try:
+        from openai.types.chat.chat_completion_message_function_tool_call import Function
+    except (ImportError, ModuleNotFoundError, AttributeError):
+        try:
+            from openai.types.chat.chat_completion_message_tool_call import Function
+        except (ImportError, ModuleNotFoundError, AttributeError):
+            pytest.skip("Could not import Function. Please check your OpenAI version. Skipping test.")
+
     openai_client.chat.completions.create(
         model="gpt-3.5-turbo",
         messages=[
             {
                 "role": "assistant",
                 "tool_calls": [
-                    ChatCompletionMessageToolCall(
+                    ChatCompletionMessageFunctionToolCall(
                         id="1",
                         type="function",
                         function=Function(
@@ -440,13 +447,21 @@ def test_chat_pydantic_based_tool_calls(
 def test_chat_pydantic_based_tool_calls_with_events_with_content(
     instrument_with_content, span_exporter, log_exporter, openai_client
 ):
+    try:
+        from openai.types.chat.chat_completion_message_function_tool_call import Function
+    except (ImportError, ModuleNotFoundError, AttributeError):
+        try:
+            from openai.types.chat.chat_completion_message_tool_call import Function
+        except (ImportError, ModuleNotFoundError, AttributeError):
+            pytest.skip("Could not import Function. Please check your OpenAI version. Skipping test.")
+
     openai_client.chat.completions.create(
         model="gpt-3.5-turbo",
         messages=[
             {
                 "role": "assistant",
                 "tool_calls": [
-                    ChatCompletionMessageToolCall(
+                    ChatCompletionMessageFunctionToolCall(
                         id="1",
                         type="function",
                         function=Function(
@@ -518,13 +533,21 @@ def test_chat_pydantic_based_tool_calls_with_events_with_content(
 def test_chat_pydantic_based_tool_calls_with_events_with_no_content(
     instrument_with_no_content, span_exporter, log_exporter, openai_client
 ):
+    try:
+        from openai.types.chat.chat_completion_message_function_tool_call import Function
+    except (ImportError, ModuleNotFoundError, AttributeError):
+        try:
+            from openai.types.chat.chat_completion_message_tool_call import Function
+        except (ImportError, ModuleNotFoundError, AttributeError):
+            pytest.skip("Could not import Function. Please check your OpenAI version. Skipping test.")
+
     openai_client.chat.completions.create(
         model="gpt-3.5-turbo",
         messages=[
             {
                 "role": "assistant",
                 "tool_calls": [
-                    ChatCompletionMessageToolCall(
+                    ChatCompletionMessageFunctionToolCall(
                         id="1",
                         type="function",
                         function=Function(
@@ -951,7 +974,6 @@ async def test_chat_async_streaming_with_events_with_no_content(
 
 
 @pytest.mark.vcr
-@pytest.mark.asyncio
 def test_with_asyncio_run(
     instrument_legacy, span_exporter, log_exporter, async_openai_client
 ):
@@ -981,7 +1003,6 @@ def test_with_asyncio_run(
 
 
 @pytest.mark.vcr
-@pytest.mark.asyncio
 def test_with_asyncio_run_with_events_with_content(
     instrument_with_content, span_exporter, log_exporter, async_openai_client
 ):
@@ -1030,7 +1051,6 @@ def test_with_asyncio_run_with_events_with_content(
 
 
 @pytest.mark.vcr
-@pytest.mark.asyncio
 def test_with_asyncio_run_with_events_with_no_content(
     instrument_with_no_content, span_exporter, log_exporter, async_openai_client
 ):

packages/opentelemetry-instrumentation-openai/tests/traces/test_chat.py

@@ -23,7 +23,7 @@ def test_responses(instrument_legacy, span_exporter: InMemorySpanExporter, opena
         span.attributes["gen_ai.prompt.0.content"] == "What is the capital of France?"
     )
     assert span.attributes["gen_ai.prompt.0.role"] == "user"
-    assert span.attributes["gen_ai.completion.0.content"] == response.output_text
+    assert span.attributes["gen_ai.completion.0.content"] == response.output[0].content[0].text
     assert span.attributes["gen_ai.completion.0.role"] == "assistant"
 
 
@@ -46,7 +46,7 @@ def test_responses_with_input_history(instrument_legacy, span_exporter: InMemory
                 "content": [
                     {
                         "type": "output_text",
-                        "text": first_response.output_text,
+                        "text": first_response.output[0].content[0].text,
                     }
                 ],
             },
@@ -69,7 +69,7 @@ def test_responses_with_input_history(instrument_legacy, span_exporter: InMemory
     assert json.loads(span.attributes["gen_ai.prompt.1.content"]) == [
         {
             "type": "output_text",
-            "text": first_response.output_text,
+            "text": first_response.output[0].content[0].text,
         }
     ]
     assert span.attributes["gen_ai.prompt.1.role"] == "assistant"
@@ -78,7 +78,7 @@ def test_responses_with_input_history(instrument_legacy, span_exporter: InMemory
         == "Can you explain why you chose that word?"
     )
     assert span.attributes["gen_ai.prompt.2.role"] == "user"
-    assert span.attributes["gen_ai.completion.0.content"] == response.output_text
+    assert span.attributes["gen_ai.completion.0.content"] == response.output[0].content[0].text
     assert span.attributes["gen_ai.completion.0.role"] == "assistant"