opentelemetry-instrumentation-openai 0.40.14__tar.gz → 0.42.0__tar.gz

{opentelemetry_instrumentation_openai-0.40.14 → opentelemetry_instrumentation_openai-0.42.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: opentelemetry-instrumentation-openai
-Version: 0.40.14
+Version: 0.42.0
 Summary: OpenTelemetry OpenAI instrumentation
 License: Apache-2.0
 Author: Gal Kleinman
@@ -17,7 +17,7 @@ Provides-Extra: instruments
 Requires-Dist: opentelemetry-api (>=1.28.0,<2.0.0)
 Requires-Dist: opentelemetry-instrumentation (>=0.50b0)
 Requires-Dist: opentelemetry-semantic-conventions (>=0.50b0)
-Requires-Dist: opentelemetry-semantic-conventions-ai (==0.4.9)
+Requires-Dist: opentelemetry-semantic-conventions-ai (==0.4.11)
 Requires-Dist: tiktoken (>=0.6.0,<1)
 Project-URL: Repository, https://github.com/traceloop/openllmetry/tree/main/packages/opentelemetry-instrumentation-openai
 Description-Content-Type: text/markdown

{opentelemetry_instrumentation_openai-0.40.14 → opentelemetry_instrumentation_openai-0.42.0}/opentelemetry/instrumentation/openai/__init__.py

@@ -1,10 +1,9 @@
 from typing import Callable, Collection, Optional
-from typing_extensions import Coroutine
 
 from opentelemetry.instrumentation.instrumentor import BaseInstrumentor
-
 from opentelemetry.instrumentation.openai.shared.config import Config
 from opentelemetry.instrumentation.openai.utils import is_openai_v1
+from typing_extensions import Coroutine
 
 _instruments = ("openai >= 0.27.0",)
 
@@ -22,6 +21,7 @@ class OpenAIInstrumentor(BaseInstrumentor):
             Callable[[str, str, str, str], Coroutine[None, None, str]]
         ] = lambda *args: "",
         enable_trace_context_propagation: bool = True,
+        use_legacy_attributes: bool = True,
     ):
         super().__init__()
         Config.enrich_assistant = enrich_assistant
@@ -30,6 +30,7 @@ class OpenAIInstrumentor(BaseInstrumentor):
         Config.get_common_metrics_attributes = get_common_metrics_attributes
         Config.upload_base64_image = upload_base64_image
         Config.enable_trace_context_propagation = enable_trace_context_propagation
+        Config.use_legacy_attributes = use_legacy_attributes
 
     def instrumentation_dependencies(self) -> Collection[str]:
         return _instruments

{opentelemetry_instrumentation_openai-0.40.14 → opentelemetry_instrumentation_openai-0.42.0}/opentelemetry/instrumentation/openai/shared/__init__.py

@@ -1,25 +1,22 @@
-import os
-import openai
 import json
-import types
 import logging
-
+import types
 from importlib.metadata import version
 
-from opentelemetry import context as context_api
-from opentelemetry.trace.propagation import set_span_in_context
-from opentelemetry.trace.propagation.tracecontext import TraceContextTextMapPropagator
-
 from opentelemetry.instrumentation.openai.shared.config import Config
-from opentelemetry.semconv._incubating.attributes.gen_ai_attributes import (
-    GEN_AI_RESPONSE_ID,
-)
-from opentelemetry.semconv_ai import SpanAttributes
 from opentelemetry.instrumentation.openai.utils import (
     dont_throw,
     is_openai_v1,
     should_record_stream_token_usage,
 )
+from opentelemetry.semconv._incubating.attributes.gen_ai_attributes import (
+    GEN_AI_RESPONSE_ID,
+)
+from opentelemetry.semconv_ai import SpanAttributes
+from opentelemetry.trace.propagation import set_span_in_context
+from opentelemetry.trace.propagation.tracecontext import TraceContextTextMapPropagator
+import openai
+import pydantic
 
 OPENAI_LLM_USAGE_TOKEN_TYPES = ["prompt_tokens", "completion_tokens"]
 PROMPT_FILTER_KEY = "prompt_filter_results"
@@ -33,12 +30,6 @@ tiktoken_encodings = {}
 logger = logging.getLogger(__name__)
 
 
-def should_send_prompts():
-    return (
-        os.getenv("TRACELOOP_TRACE_CONTENT") or "true"
-    ).lower() == "true" or context_api.get_value("override_enable_content_tracing")
-
-
 def _set_span_attribute(span, name, value):
     if value is None or value == "":
         return
@@ -113,13 +104,23 @@ def set_tools_attributes(span, tools):
         )
 
 
-def _set_request_attributes(span, kwargs):
+def _set_request_attributes(span, kwargs, instance=None):
     if not span.is_recording():
         return
 
     _set_api_attributes(span)
-    _set_span_attribute(span, SpanAttributes.LLM_SYSTEM, "OpenAI")
-    _set_span_attribute(span, SpanAttributes.LLM_REQUEST_MODEL, kwargs.get("model"))
+
+    base_url = _get_openai_base_url(instance) if instance else ""
+    vendor = _get_vendor_from_url(base_url)
+    _set_span_attribute(span, SpanAttributes.LLM_SYSTEM, vendor)
+
+    model = kwargs.get("model")
+    if vendor == "AWS" and model and "." in model:
+        model = _cross_region_check(model)
+    elif vendor == "OpenRouter":
+        model = _extract_model_name_from_provider_format(model)
+
+    _set_span_attribute(span, SpanAttributes.LLM_REQUEST_MODEL, model)
     _set_span_attribute(
         span, SpanAttributes.LLM_REQUEST_MAX_TOKENS, kwargs.get("max_tokens")
     )
@@ -143,6 +144,49 @@ def _set_request_attributes(span, kwargs):
     _set_span_attribute(
         span, SpanAttributes.LLM_IS_STREAMING, kwargs.get("stream") or False
     )
+    if response_format := kwargs.get("response_format"):
+        # backward-compatible check for
+        # openai.types.shared_params.response_format_json_schema.ResponseFormatJSONSchema
+        if (
+            isinstance(response_format, dict)
+            and response_format.get("type") == "json_schema"
+            and response_format.get("json_schema")
+        ):
+            schema = dict(response_format.get("json_schema")).get("schema")
+            if schema:
+                _set_span_attribute(
+                    span,
+                    SpanAttributes.LLM_REQUEST_STRUCTURED_OUTPUT_SCHEMA,
+                    json.dumps(schema),
+                )
+        elif (
+            isinstance(response_format, pydantic.BaseModel)
+            or (
+                hasattr(response_format, "model_json_schema")
+                and callable(response_format.model_json_schema)
+            )
+        ):
+            _set_span_attribute(
+                span,
+                SpanAttributes.LLM_REQUEST_STRUCTURED_OUTPUT_SCHEMA,
+                json.dumps(response_format.model_json_schema()),
+            )
+        else:
+            schema = None
+            try:
+                schema = json.dumps(pydantic.TypeAdapter(response_format).json_schema())
+            except Exception:
+                try:
+                    schema = json.dumps(response_format)
+                except Exception:
+                    pass
+
+            if schema:
+                _set_span_attribute(
+                    span,
+                    SpanAttributes.LLM_REQUEST_STRUCTURED_OUTPUT_SCHEMA,
+                    schema,
+                )
 
 
 @dont_throw
@@ -158,7 +202,10 @@ def _set_response_attributes(span, response):
         )
         return
 
-    _set_span_attribute(span, SpanAttributes.LLM_RESPONSE_MODEL, response.get("model"))
+    response_model = response.get("model")
+    if response_model:
+        response_model = _extract_model_name_from_provider_format(response_model)
+    _set_span_attribute(span, SpanAttributes.LLM_RESPONSE_MODEL, response_model)
     _set_span_attribute(span, GEN_AI_RESPONSE_ID, response.get("id"))
 
     _set_span_attribute(
@@ -187,7 +234,9 @@ def _set_response_attributes(span, response):
     )
     prompt_tokens_details = dict(usage.get("prompt_tokens_details", {}))
     _set_span_attribute(
-        span, SpanAttributes.LLM_USAGE_CACHE_READ_INPUT_TOKENS, prompt_tokens_details.get("cached_tokens", 0)
+        span,
+        SpanAttributes.LLM_USAGE_CACHE_READ_INPUT_TOKENS,
+        prompt_tokens_details.get("cached_tokens", 0),
     )
     return
 
@@ -206,17 +255,17 @@ def _set_span_stream_usage(span, prompt_tokens, completion_tokens):
     if not span.is_recording():
         return
 
-    if type(completion_tokens) is int and completion_tokens >= 0:
+    if isinstance(completion_tokens, int) and completion_tokens >= 0:
         _set_span_attribute(
             span, SpanAttributes.LLM_USAGE_COMPLETION_TOKENS, completion_tokens
         )
 
-    if type(prompt_tokens) is int and prompt_tokens >= 0:
+    if isinstance(prompt_tokens, int) and prompt_tokens >= 0:
         _set_span_attribute(span, SpanAttributes.LLM_USAGE_PROMPT_TOKENS, prompt_tokens)
 
     if (
-        type(prompt_tokens) is int
-        and type(completion_tokens) is int
+        isinstance(prompt_tokens, int)
+        and isinstance(completion_tokens, int)
         and completion_tokens + prompt_tokens >= 0
     ):
         _set_span_attribute(
@@ -235,6 +284,53 @@ def _get_openai_base_url(instance):
     return ""
 
 
+def _get_vendor_from_url(base_url):
+    if not base_url:
+        return "openai"
+
+    if "openai.azure.com" in base_url:
+        return "Azure"
+    elif "amazonaws.com" in base_url or "bedrock" in base_url:
+        return "AWS"
+    elif "googleapis.com" in base_url or "vertex" in base_url:
+        return "Google"
+    elif "openrouter.ai" in base_url:
+        return "OpenRouter"
+
+    return "openai"
+
+
+def _cross_region_check(value):
+    if not value or "." not in value:
+        return value
+
+    prefixes = ["us", "us-gov", "eu", "apac"]
+    if any(value.startswith(prefix + ".") for prefix in prefixes):
+        parts = value.split(".")
+        if len(parts) > 2:
+            return parts[2]
+        else:
+            return value
+    else:
+        vendor, model = value.split(".", 1)
+        return model
+
+
+def _extract_model_name_from_provider_format(model_name):
+    """
+    Extract model name from provider/model format.
+    E.g., 'openai/gpt-4o' -> 'gpt-4o', 'anthropic/claude-3-sonnet' -> 'claude-3-sonnet'
+    """
+    if not model_name:
+        return model_name
+
+    if "/" in model_name:
+        parts = model_name.split("/")
+        return parts[-1]  # Return the last part (actual model name)
+
+    return model_name
+
+
 def is_streaming_response(response):
     if is_openai_v1():
         return isinstance(response, openai.Stream) or isinstance(
@@ -296,10 +392,11 @@ def metric_shared_attributes(
     response_model: str, operation: str, server_address: str, is_streaming: bool = False
 ):
     attributes = Config.get_common_metrics_attributes()
+    vendor = _get_vendor_from_url(server_address)
 
     return {
         **attributes,
-        SpanAttributes.LLM_SYSTEM: "openai",
+        SpanAttributes.LLM_SYSTEM: vendor,
         SpanAttributes.LLM_RESPONSE_MODEL: response_model,
         "gen_ai.operation.name": operation,
         "server.address": server_address,

{opentelemetry_instrumentation_openai-0.40.14 → opentelemetry_instrumentation_openai-0.42.0}/opentelemetry/instrumentation/openai/shared/chat_wrappers.py

@@ -2,47 +2,57 @@ import copy
 import json
 import logging
 import time
-from opentelemetry.instrumentation.openai.shared.config import Config
-from wrapt import ObjectProxy
-
+from functools import singledispatch
+from typing import List, Optional, Union
 
 from opentelemetry import context as context_api
-from opentelemetry.metrics import Counter, Histogram
-from opentelemetry.semconv_ai import (
-    SUPPRESS_LANGUAGE_MODEL_INSTRUMENTATION_KEY,
-    SpanAttributes,
-    LLMRequestTypeValues,
-)
-
-from opentelemetry.instrumentation.utils import _SUPPRESS_INSTRUMENTATION_KEY
-from opentelemetry.instrumentation.openai.utils import (
-    _with_chat_telemetry_wrapper,
-    dont_throw,
-    run_async,
-)
 from opentelemetry.instrumentation.openai.shared import (
-    metric_shared_attributes,
+    OPENAI_LLM_USAGE_TOKEN_TYPES,
+    _get_openai_base_url,
     _set_client_attributes,
+    _set_functions_attributes,
     _set_request_attributes,
+    _set_response_attributes,
     _set_span_attribute,
-    _set_functions_attributes,
+    _set_span_stream_usage,
     _token_type,
-    set_tools_attributes,
-    _set_response_attributes,
+    get_token_count_from_string,
     is_streaming_response,
-    should_send_prompts,
+    metric_shared_attributes,
     model_as_dict,
-    _get_openai_base_url,
-    OPENAI_LLM_USAGE_TOKEN_TYPES,
-    should_record_stream_token_usage,
-    get_token_count_from_string,
-    _set_span_stream_usage,
     propagate_trace_context,
+    set_tools_attributes,
+    should_record_stream_token_usage,
+)
+from opentelemetry.instrumentation.openai.shared.config import Config
+from opentelemetry.instrumentation.openai.shared.event_emitter import emit_event
+from opentelemetry.instrumentation.openai.shared.event_models import (
+    ChoiceEvent,
+    MessageEvent,
+    ToolCall,
+)
+from opentelemetry.instrumentation.openai.utils import (
+    _with_chat_telemetry_wrapper,
+    dont_throw,
+    is_openai_v1,
+    run_async,
+    should_emit_events,
+    should_send_prompts,
+)
+from opentelemetry.instrumentation.utils import _SUPPRESS_INSTRUMENTATION_KEY
+from opentelemetry.metrics import Counter, Histogram
+from opentelemetry.semconv.attributes.error_attributes import ERROR_TYPE
+from opentelemetry.semconv_ai import (
+    SUPPRESS_LANGUAGE_MODEL_INSTRUMENTATION_KEY,
+    LLMRequestTypeValues,
+    SpanAttributes,
 )
 from opentelemetry.trace import SpanKind, Tracer
 from opentelemetry.trace.status import Status, StatusCode
+from wrapt import ObjectProxy
 
-from opentelemetry.instrumentation.openai.utils import is_openai_v1
+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"
@@ -80,7 +90,6 @@ def chat_wrapper(
     )
 
     run_async(_handle_request(span, kwargs, instance))
-
     try:
         start_time = time.time()
         response = wrapped(*args, **kwargs)
@@ -98,10 +107,12 @@ def chat_wrapper(
         if exception_counter:
             exception_counter.add(1, attributes=attributes)
 
+        span.set_attribute(ERROR_TYPE, e.__class__.__name__)
+        span.record_exception(e)
         span.set_status(Status(StatusCode.ERROR, str(e)))
         span.end()
 
-        raise e
+        raise
 
     if is_streaming_response(response):
         # span will be closed after the generator is done
@@ -143,6 +154,7 @@ def chat_wrapper(
         duration_histogram,
         duration,
     )
+
     span.end()
 
     return response
@@ -172,6 +184,7 @@ async def achat_wrapper(
         kind=SpanKind.CLIENT,
         attributes={SpanAttributes.LLM_REQUEST_TYPE: LLM_REQUEST_TYPE.value},
     )
+
     await _handle_request(span, kwargs, instance)
 
     try:
@@ -193,10 +206,12 @@ async def achat_wrapper(
         if exception_counter:
             exception_counter.add(1, attributes=attributes)
 
+        span.set_attribute(ERROR_TYPE, e.__class__.__name__)
+        span.record_exception(e)
         span.set_status(Status(StatusCode.ERROR, str(e)))
         span.end()
 
-        raise e
+        raise
 
     if is_streaming_response(response):
         # span will be closed after the generator is done
@@ -238,6 +253,7 @@ async def achat_wrapper(
         duration_histogram,
         duration,
     )
+
     span.end()
 
     return response
@@ -245,14 +261,24 @@ async def achat_wrapper(
 
 @dont_throw
 async def _handle_request(span, kwargs, instance):
-    _set_request_attributes(span, kwargs)
+    _set_request_attributes(span, kwargs, instance)
     _set_client_attributes(span, instance)
-    if should_send_prompts():
-        await _set_prompts(span, kwargs.get("messages"))
-        if kwargs.get("functions"):
-            _set_functions_attributes(span, kwargs.get("functions"))
-        elif kwargs.get("tools"):
-            set_tools_attributes(span, kwargs.get("tools"))
+    if should_emit_events():
+        for message in kwargs.get("messages", []):
+            emit_event(
+                MessageEvent(
+                    content=message.get("content"),
+                    role=message.get("role"),
+                    tool_calls=_parse_tool_calls(message.get("tool_calls", None)),
+                )
+            )
+    else:
+        if should_send_prompts():
+            await _set_prompts(span, kwargs.get("messages"))
+            if kwargs.get("functions"):
+                _set_functions_attributes(span, kwargs.get("functions"))
+            elif kwargs.get("tools"):
+                set_tools_attributes(span, kwargs.get("tools"))
     if Config.enable_trace_context_propagation:
         propagate_trace_context(span, kwargs)
 
@@ -285,8 +311,13 @@ def _handle_response(
     # span attributes
     _set_response_attributes(span, response_dict)
 
-    if should_send_prompts():
-        _set_completions(span, response_dict.get("choices"))
+    if should_emit_events():
+        if response.choices is not None:
+            for choice in response.choices:
+                emit_event(_parse_choice_event(choice))
+    else:
+        if should_send_prompts():
+            _set_completions(span, response_dict.get("choices"))
 
     return response
 
@@ -528,14 +559,14 @@ def _set_streaming_token_metrics(
 
     # metrics record
     if token_counter:
-        if type(prompt_usage) is int and prompt_usage >= 0:
+        if isinstance(prompt_usage, int) and prompt_usage >= 0:
             attributes_with_token_type = {
                 **shared_attributes,
                 SpanAttributes.LLM_TOKEN_TYPE: "input",
             }
             token_counter.record(prompt_usage, attributes=attributes_with_token_type)
 
-        if type(completion_usage) is int and completion_usage >= 0:
+        if isinstance(completion_usage, int) and completion_usage >= 0:
             attributes_with_token_type = {
                 **shared_attributes,
                 SpanAttributes.LLM_TOKEN_TYPE: "output",
@@ -609,8 +640,8 @@ class ChatStream(ObjectProxy):
             chunk = self.__wrapped__.__next__()
         except Exception as e:
             if isinstance(e, StopIteration):
-                self._close_span()
-            raise e
+                self._process_complete_response()
+            raise
         else:
             self._process_item(chunk)
             return chunk
@@ -620,8 +651,8 @@ class ChatStream(ObjectProxy):
             chunk = await self.__wrapped__.__anext__()
         except Exception as e:
             if isinstance(e, StopAsyncIteration):
-                self._close_span()
-            raise e
+                self._process_complete_response()
+            raise
         else:
             self._process_item(chunk)
             return chunk
@@ -650,7 +681,7 @@ class ChatStream(ObjectProxy):
         )
 
     @dont_throw
-    def _close_span(self):
+    def _process_complete_response(self):
         _set_streaming_token_metrics(
             self._request_kwargs,
             self._complete_response,
@@ -683,9 +714,12 @@ class ChatStream(ObjectProxy):
             )
 
         _set_response_attributes(self._span, self._complete_response)
-
-        if should_send_prompts():
-            _set_completions(self._span, self._complete_response.get("choices"))
+        if should_emit_events():
+            for choice in self._complete_response.get("choices", []):
+                emit_event(_parse_choice_event(choice))
+        else:
+            if should_send_prompts():
+                _set_completions(self._span, self._complete_response.get("choices"))
 
         self._span.set_status(Status(StatusCode.OK))
         self._span.end()
@@ -753,9 +787,12 @@ def _build_from_streaming_response(
         streaming_time_to_generate.record(time.time() - time_of_first_token)
 
     _set_response_attributes(span, complete_response)
-
-    if should_send_prompts():
-        _set_completions(span, complete_response.get("choices"))
+    if should_emit_events():
+        for choice in complete_response.get("choices", []):
+            emit_event(_parse_choice_event(choice))
+    else:
+        if should_send_prompts():
+            _set_completions(span, complete_response.get("choices"))
 
     span.set_status(Status(StatusCode.OK))
     span.end()
@@ -820,14 +857,113 @@ async def _abuild_from_streaming_response(
         streaming_time_to_generate.record(time.time() - time_of_first_token)
 
     _set_response_attributes(span, complete_response)
-
-    if should_send_prompts():
-        _set_completions(span, complete_response.get("choices"))
+    if should_emit_events():
+        for choice in complete_response.get("choices", []):
+            emit_event(_parse_choice_event(choice))
+    else:
+        if should_send_prompts():
+            _set_completions(span, complete_response.get("choices"))
 
     span.set_status(Status(StatusCode.OK))
     span.end()
 
 
+def _parse_tool_calls(
+    tool_calls: Optional[List[Union[dict, ChatCompletionMessageToolCall]]],
+) -> Union[List[ToolCall], None]:
+    """
+    Util to correctly parse the tool calls data from the OpenAI API to this module's
+    standard `ToolCall`.
+    """
+    if tool_calls is None:
+        return tool_calls
+
+    result = []
+
+    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):
+            tool_call_data = tool_call.model_dump()
+        elif isinstance(tool_call, FunctionCall):
+            function_call = tool_call.model_dump()
+            tool_call_data = ToolCall(
+                id="",
+                function={
+                    "name": function_call.get("name"),
+                    "arguments": function_call.get("arguments"),
+                },
+                type="function",
+            )
+
+        result.append(tool_call_data)
+    return result
+
+
+@singledispatch
+def _parse_choice_event(choice) -> ChoiceEvent:
+    has_message = choice.message is not None
+    has_finish_reason = choice.finish_reason is not None
+    has_tool_calls = has_message and choice.message.tool_calls
+    has_function_call = has_message and choice.message.function_call
+
+    content = choice.message.content if has_message else None
+    role = choice.message.role if has_message else "unknown"
+    finish_reason = choice.finish_reason if has_finish_reason else "unknown"
+
+    if has_tool_calls and has_function_call:
+        tool_calls = choice.message.tool_calls + [choice.message.function_call]
+    elif has_tool_calls:
+        tool_calls = choice.message.tool_calls
+    elif has_function_call:
+        tool_calls = [choice.message.function_call]
+    else:
+        tool_calls = None
+
+    return ChoiceEvent(
+        index=choice.index,
+        message={"content": content, "role": role},
+        finish_reason=finish_reason,
+        tool_calls=_parse_tool_calls(tool_calls),
+    )
+
+
+@_parse_choice_event.register
+def _(choice: dict) -> ChoiceEvent:
+    message = choice.get("message")
+    has_message = message is not None
+    has_finish_reason = choice.get("finish_reason") is not None
+    has_tool_calls = has_message and message.get("tool_calls")
+    has_function_call = has_message and message.get("function_call")
+
+    content = choice.get("message").get("content", "") if has_message else None
+    role = choice.get("message").get("role") if has_message else "unknown"
+    finish_reason = choice.get("finish_reason") if has_finish_reason else "unknown"
+
+    if has_tool_calls and has_function_call:
+        tool_calls = message.get("tool_calls") + [message.get("function_call")]
+    elif has_tool_calls:
+        tool_calls = message.get("tool_calls")
+    elif has_function_call:
+        tool_calls = [message.get("function_call")]
+    else:
+        tool_calls = None
+
+    if tool_calls is not None:
+        for tool_call in tool_calls:
+            tool_call["type"] = "function"
+
+    return ChoiceEvent(
+        index=choice.get("index"),
+        message={"content": content, "role": role},
+        finish_reason=finish_reason,
+        tool_calls=tool_calls,
+    )
+
+
 def _accumulate_stream_items(item, complete_response):
     if is_openai_v1():
         item = model_as_dict(item)