Vertex response gen_ai.choice events

Author: aabmass

instrumentation-genai/opentelemetry-instrumentation-vertexai/src/opentelemetry/instrumentation/vertexai/events.py

@@ -20,6 +20,11 @@
 schematized in YAML and the Weaver tool supports it.
 """
 
+from __future__ import annotations
+
+from dataclasses import asdict, dataclass
+from typing import Literal
+
 from opentelemetry._events import Event
 from opentelemetry.semconv._incubating.attributes import gen_ai_attributes
 from opentelemetry.util.types import AnyValue
@@ -89,3 +94,42 @@ def system_event(
         },
         body=body,
     )
+
+
+@dataclass
+class ChoiceMessage:
+    """The message field for a gen_ai.choice event"""
+
+    content: AnyValue = None
+    role: str = "assistant"
+
+
+FinishReason = Literal[
+    "content_filter", "error", "length", "stop", "tool_calls"
+]
+
+
+# TODO add tool calls
+# https://github.com/open-telemetry/opentelemetry-python-contrib/issues/3216
+def choice_event(
+    *,
+    finish_reason: FinishReason | str,
+    index: int,
+    message: ChoiceMessage,
+) -> Event:
+    """Creates a choice event, which describes the Gen AI response message.
+    https://github.com/open-telemetry/semantic-conventions/blob/v1.28.0/docs/gen-ai/gen-ai-events.md#event-gen_aichoice
+    """
+    body: dict[str, AnyValue] = {
+        "finish_reason": finish_reason,
+        "index": index,
+        "message": asdict(message),
+    }
+
+    return Event(
+        name="gen_ai.choice",
+        attributes={
+            gen_ai_attributes.GEN_AI_SYSTEM: gen_ai_attributes.GenAiSystemValues.VERTEX_AI.value,
+        },
+        body=body,
+    )

instrumentation-genai/opentelemetry-instrumentation-vertexai/src/opentelemetry/instrumentation/vertexai/patch.py

@@ -29,6 +29,7 @@
     get_server_attributes,
     get_span_name,
     request_to_events,
+    response_to_events,
 )
 from opentelemetry.trace import SpanKind, Tracer
 
@@ -131,10 +132,11 @@ def traced_method(
 
             if span.is_recording():
                 span.set_attributes(get_genai_response_attributes(response))
-            # TODO: add response attributes and events
-            # _set_response_attributes(
-            #     span, result, event_logger, capture_content
-            # )
+            for event in response_to_events(
+                response=response, capture_content=capture_content
+            ):
+                event_logger.emit(event)
+
             return response
 
     return traced_method

instrumentation-genai/opentelemetry-instrumentation-vertexai/src/opentelemetry/instrumentation/vertexai/utils.py

@@ -28,7 +28,10 @@
 
 from opentelemetry._events import Event
 from opentelemetry.instrumentation.vertexai.events import (
+    ChoiceMessage,
+    FinishReason,
     assistant_event,
+    choice_event,
     system_event,
     user_event,
 )
@@ -55,6 +58,9 @@
     )
 
 
+_MODEL = "model"
+
+
 @dataclass(frozen=True)
 class GenerateContentParams:
     model: str
@@ -204,7 +210,7 @@ def request_to_events(
 
     for content in params.contents or []:
         # Assistant message
-        if content.role == "model":
+        if content.role == _MODEL:
             request_content = _parts_to_any_value(
                 capture_content=capture_content, parts=content.parts
             )
@@ -218,6 +224,27 @@ def request_to_events(
             yield user_event(role=content.role, content=request_content)
 
 
+def response_to_events(
+    *,
+    response: prediction_service.GenerateContentResponse
+    | prediction_service_v1beta1.GenerateContentResponse,
+    capture_content: bool,
+) -> Iterable[Event]:
+    for candidate in response.candidates:
+        yield choice_event(
+            finish_reason=_map_finish_reason(candidate.finish_reason),
+            index=candidate.index,
+            # default to "model" since Vertex uses that instead of assistant
+            message=ChoiceMessage(
+                role=candidate.content.role or _MODEL,
+                content=_parts_to_any_value(
+                    capture_content=capture_content,
+                    parts=candidate.content.parts,
+                ),
+            ),
+        )
+
+
 def _parts_to_any_value(
     *,
     capture_content: bool,
@@ -230,3 +257,26 @@ def _parts_to_any_value(
         cast("dict[str, AnyValue]", type(part).to_dict(part))  # type: ignore[reportUnknownMemberType]
         for part in parts
     ]
+
+
+def _map_finish_reason(
+    finish_reason: content.Candidate.FinishReason
+    | content_v1beta1.Candidate.FinishReason,
+) -> FinishReason | str:
+    EnumType = type(finish_reason)
+    if (
+        finish_reason is EnumType.FINISH_REASON_UNSPECIFIED
+        or finish_reason is EnumType.OTHER
+    ):
+        return "error"
+    if finish_reason is EnumType.STOP:
+        return "stop"
+    if finish_reason is EnumType.MAX_TOKENS:
+        return "length"
+
+    # There are a lot of specific enum values from Vertex that would map to "content_filter".
+    # I'm worried trying to map the enum obfuscates the telemetry because 1) it over
+    # generalizes and 2) half of the values are from the OTel enum and others from the vertex
+    # enum. See for reference
+    # https://github.com/googleapis/python-aiplatform/blob/c5023698c7068e2f84523f91b824641c9ef2d694/google/cloud/aiplatform_v1/types/content.py#L786-L822
+    return finish_reason.name.lower()