simplify and avoid circular deps

Author: lmolkova

util/opentelemetry-util-genai/src/opentelemetry/util/genai/embedding_invocation.py

@@ -14,17 +14,17 @@
 
 from __future__ import annotations
 
-from typing import TYPE_CHECKING, Any
+from typing import Any
 
+from opentelemetry._logs import Logger
 from opentelemetry.semconv._incubating.attributes import (
     gen_ai_attributes as GenAI,
 )
 from opentelemetry.semconv.attributes import server_attributes
-from opentelemetry.trace import SpanKind
+from opentelemetry.trace import SpanKind, Tracer
+from opentelemetry.util.genai.metrics import InvocationMetricsRecorder
 from opentelemetry.util.genai.types import Error, GenAIInvocation
-
-if TYPE_CHECKING:
-    from opentelemetry.util.genai.handler import TelemetryHandler
+from opentelemetry.util.types import AttributeValue
 
 
 class EmbeddingInvocation(GenAIInvocation):
@@ -34,13 +34,11 @@ class EmbeddingInvocation(GenAIInvocation):
     context manager rather than constructing this directly.
     """
 
-    @property
-    def operation_name(self) -> str:
-        return GenAI.GenAiOperationNameValues.EMBEDDINGS.value
-
     def __init__(
         self,
-        handler: TelemetryHandler,
+        tracer: Tracer,
+        metrics_recorder: InvocationMetricsRecorder,
+        logger: Logger,
         provider: str,
         *,
         request_model: str | None = None,
@@ -55,8 +53,13 @@ def __init__(
     ) -> None:
         """Use handler.start_embedding(provider) or handler.embedding(provider) instead of calling this directly."""
         super().__init__(
-            handler, attributes=attributes, metric_attributes=metric_attributes
+            tracer,
+            metrics_recorder,
+            logger,
+            attributes=attributes,
+            metric_attributes=metric_attributes,
         )
+        self._operation_name = GenAI.GenAiOperationNameValues.EMBEDDINGS.value
         self.provider = provider  # e.g., azure.ai.openai, openai, aws.bedrock
         self.request_model = request_model
         self.server_address = server_address
@@ -68,12 +71,32 @@ def __init__(
         self.dimension_count = dimension_count
         self.response_model_name = response_model_name
         self._span_name = (
-            f"{self.operation_name} {request_model}"
+            f"{self._operation_name} {request_model}"
             if request_model
-            else self.operation_name
+            else self._operation_name
         )
         self._span_kind = SpanKind.CLIENT
-        handler._start(self)
+        self._start()
+
+    def _get_metric_attributes(self) -> dict[str, Any]:
+        optional_attrs = (
+            (GenAI.GEN_AI_PROVIDER_NAME, self.provider),
+            (GenAI.GEN_AI_REQUEST_MODEL, self.request_model),
+            (GenAI.GEN_AI_RESPONSE_MODEL, self.response_model_name),
+            (server_attributes.SERVER_ADDRESS, self.server_address),
+            (server_attributes.SERVER_PORT, self.server_port),
+        )
+        attrs: dict[str, AttributeValue] = {
+            GenAI.GEN_AI_OPERATION_NAME: self._operation_name,
+            **{k: v for k, v in optional_attrs if v is not None},
+        }
+        attrs.update(self.metric_attributes)
+        return attrs
+
+    def _get_metric_token_counts(self) -> dict[str, int]:
+        if self.input_tokens is not None:
+            return {GenAI.GenAiTokenTypeValues.INPUT.value: self.input_tokens}
+        return {}
 
     def _apply_finish(self, error: Error | None = None) -> None:
         optional_attrs = (
@@ -87,7 +110,7 @@ def _apply_finish(self, error: Error | None = None) -> None:
             (GenAI.GEN_AI_USAGE_INPUT_TOKENS, self.input_tokens),
         )
         attributes: dict[str, Any] = {
-            GenAI.GEN_AI_OPERATION_NAME: self.operation_name,
+            GenAI.GEN_AI_OPERATION_NAME: self._operation_name,
             **{
                 key: value
                 for key, value in optional_attrs

util/opentelemetry-util-genai/src/opentelemetry/util/genai/handler.py

@@ -48,13 +48,11 @@
 
 from __future__ import annotations
 
-import timeit
 from contextlib import contextmanager
-from typing import Iterator, TypeVar
+from typing import Iterator
 
 from typing_extensions import deprecated
 
-from opentelemetry import context as otel_context
 from opentelemetry._logs import (
     LoggerProvider,
     get_logger,
@@ -64,7 +62,6 @@
 from opentelemetry.trace import (
     TracerProvider,
     get_tracer,
-    set_span_in_context,
 )
 from opentelemetry.util.genai.embedding_invocation import EmbeddingInvocation
 from opentelemetry.util.genai.inference_invocation import (
@@ -73,30 +70,11 @@
 )
 from opentelemetry.util.genai.metrics import InvocationMetricsRecorder
 from opentelemetry.util.genai.tool_invocation import ToolInvocation
-from opentelemetry.util.genai.types import (
-    Error,
-    GenAIInvocation,
-)
+from opentelemetry.util.genai.types import Error
 from opentelemetry.util.genai.version import __version__
 from opentelemetry.util.genai.workflow_invocation import WorkflowInvocation
 
 
-def _safe_detach(invocation: GenAIInvocation) -> None:
-    """Detach the context token if still present, as a safety net."""
-    if invocation._context_token is not None:
-        try:
-            otel_context.detach(invocation._context_token)
-        except Exception:  # pylint: disable=broad-except
-            pass
-    try:
-        invocation.span.end()
-    except Exception:  # pylint: disable=broad-except
-        pass
-
-
-_T = TypeVar("_T", bound=GenAIInvocation)
-
-
 class TelemetryHandler:
     """
     High-level handler managing GenAI invocation lifecycles and emitting
@@ -127,40 +105,6 @@ def __init__(
             schema_url=schema_url,
         )
 
-    def _start(self, invocation: _T) -> _T:
-        """Start a GenAI invocation and create a pending span entry."""
-
-        invocation._handler = self
-
-        invocation.span = self._tracer.start_span(
-            name=invocation._span_name,
-            kind=invocation._span_kind,
-        )
-
-        # Record a monotonic start timestamp (seconds) for duration
-        # calculation using timeit.default_timer.
-        invocation._monotonic_start_s = timeit.default_timer()
-        invocation._context_token = otel_context.attach(
-            set_span_in_context(invocation.span)
-        )
-
-        return invocation
-
-    def _finish(  # pylint: disable=no-self-use
-        self, invocation: _T, error: Error | None = None
-    ) -> _T:
-        """Finalize a GenAI invocation and end its span."""
-        if invocation._context_token is None:
-            # TODO: Provide feedback that this invocation was not started
-            return invocation
-
-        try:
-            invocation._apply_finish(error)
-        finally:
-            # Detach context and end span even if finishing fails
-            _safe_detach(invocation)
-        return invocation
-
     # New-style factory methods: construct + start in one call, handler stored on invocation
 
     def start_inference(
@@ -177,7 +121,9 @@ def start_inference(
         returned invocation, then call invocation.stop() or invocation.fail().
         """
         return InferenceInvocation(
-            self,
+            self._tracer,
+            self._metrics_recorder,
+            self._logger,
             provider,
             request_model=request_model,
             server_address=server_address,
@@ -194,9 +140,12 @@ def start_llm(self, invocation: LLMInvocation) -> LLMInvocation:  # pyright: ign
             Use ``handler.start_inference()`` instead.
         """
         if invocation._context_token is not None:
-            # Already started (e.g. via InferenceInvocation.__init__)
+            # Already started (e.g. tracer passed to LLMInvocation.__init__)
             return invocation
-        return self._start(invocation)
+        invocation._start_with_handler(
+            self._tracer, self._metrics_recorder, self._logger
+        )
+        return invocation
 
     def start_embedding(
         self,
@@ -212,7 +161,9 @@ def start_embedding(
         invocation, then call invocation.stop() or invocation.fail().
         """
         return EmbeddingInvocation(
-            self,
+            self._tracer,
+            self._metrics_recorder,
+            self._logger,
             provider,
             request_model=request_model,
             server_address=server_address,
@@ -234,7 +185,9 @@ def start_tool(
         invocation.stop() or invocation.fail().
         """
         return ToolInvocation(
-            self,
+            self._tracer,
+            self._metrics_recorder,
+            self._logger,
             name,
             arguments=arguments,
             tool_call_id=tool_call_id,
@@ -251,7 +204,9 @@ def start_workflow(
         Set remaining attributes on the returned invocation, then call
         invocation.stop() or invocation.fail().
         """
-        return WorkflowInvocation(self, name)
+        return WorkflowInvocation(
+            self._tracer, self._metrics_recorder, self._logger, name
+        )
 
     @deprecated(
         "handler.stop_llm() is deprecated. Use invocation.stop() instead."
@@ -262,7 +217,8 @@ def stop_llm(self, invocation: LLMInvocation) -> LLMInvocation:  # pyright: igno
         .. deprecated::
             Use ``invocation.stop()`` instead.
         """
-        return self._finish(invocation)
+        invocation.stop()
+        return invocation
 
     @deprecated(
         "handler.fail_llm() is deprecated. Use invocation.fail(error) instead."
@@ -277,7 +233,8 @@ def fail_llm(
         .. deprecated::
             Use ``invocation.fail(error)`` instead.
         """
-        return self._finish(invocation, error)
+        invocation.fail(error)
+        return invocation
 
     @contextmanager
     def inference(