WIP gen_ai chat refactor

Author: keith-decker

util/opentelemetry-util-genai/README.rst

@@ -6,6 +6,24 @@ The GenAI Utils package will include boilerplate and helpers to standardize inst
 This package will provide APIs and decorators to minimize the work needed to instrument genai libraries, 
 while providing standardization for generating both types of otel, "spans and metrics" and "spans, metrics and events"
 
+This package provides these span attributes.
+-> gen_ai.operation.name: Str(chat)
+-> gen_ai.system: Str(ChatOpenAI)
+-> gen_ai.request.model: Str(gpt-3.5-turbo)
+-> gen_ai.request.top_p: Double(0.9)
+-> gen_ai.request.frequency_penalty: Double(0.5)
+-> gen_ai.request.presence_penalty: Double(0.5)
+-> gen_ai.request.stop_sequences: Slice(["\n","Human:","AI:"])
+-> gen_ai.request.seed: Int(100)
+-> gen_ai.request.max_tokens: Int(100)
+-> gen_ai.provider.name: Str(openai)
+-> gen_ai.request.temperature: Double(0.1)
+-> gen_ai.response.finish_reasons: Slice(["stop"])
+-> gen_ai.response.model: Str(gpt-3.5-turbo-0125)
+-> gen_ai.response.id: Str(chatcmpl-Bz8yrvPnydD9pObv625n2CGBPHS13)
+-> gen_ai.usage.input_tokens: Int(24)
+-> gen_ai.usage.output_tokens: Int(7)
+
 Installation
 ------------
 

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

@@ -42,23 +42,23 @@ def __init__(self, emitter_type_full: bool = True, **kwargs):
             __name__,
             __version__,
             tracer_provider,
-            schema_url=Schemas.V1_28_0.value,
+            schema_url=Schemas.V1_36_0.value,
         )
 
         meter_provider = kwargs.get("meter_provider")
         self._meter = get_meter(
             __name__,
             __version__,
             meter_provider,
-            schema_url=Schemas.V1_28_0.value,
+            schema_url=Schemas.V1_36_0.value,
         )
 
         event_logger_provider = kwargs.get("event_logger_provider")
         self._event_logger = get_event_logger(
             __name__,
             __version__,
             event_logger_provider=event_logger_provider,
-            schema_url=Schemas.V1_28_0.value,
+            schema_url=Schemas.V1_36_0.value,
         )
 
         self._emitter = (

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

@@ -7,6 +7,23 @@ class Message:
     type: str
     name: str
 
+    def _to_part_dict(self):
+        """Convert the message to a dictionary suitable for OpenTelemetry semconvs.
+
+        Ref: https://github.com/open-telemetry/semantic-conventions/blob/main/docs/registry/attributes/gen-ai.md#gen-ai-input-messages
+        """
+
+        # Support tool_call and tool_call response
+        return {
+            "role": self.type,
+            "parts": [
+                {
+                    "content": self.content,
+                    "type": "text",
+                }
+            ],
+        }
+
 
 @dataclass
 class ChatGeneration:

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

@@ -34,6 +34,7 @@
     use_span,
 )
 from opentelemetry.trace.status import Status, StatusCode
+from opentelemetry.util.types import Attributes
 
 from .data import Error
 from .instruments import Instruments
@@ -59,8 +60,11 @@ def _get_property_value(obj, property_name) -> object:
 
 
 def _message_to_event(message, system, framework) -> Optional[Event]:
+    # TODO: Convert to logs.
     content = _get_property_value(message, "content")
     if content:
+        # update this to event.gen_ai.client.inference.operation.details: https://github.com/open-telemetry/semantic-conventions/blob/main/docs/gen-ai/gen-ai-events.md
+
         message_type = _get_property_value(message, "type")
         message_type = "user" if message_type == "human" else message_type
         body = {"content": content}
@@ -80,6 +84,7 @@ def _message_to_event(message, system, framework) -> Optional[Event]:
 def _chat_generation_to_event(
     chat_generation, index, system, framework
 ) -> Optional[Event]:
+    # TODO: Convert to logs.
     if chat_generation.content:
         attributes = {
             # TODO: add below to opentelemetry.semconv._incubating.attributes.gen_ai_attributes
@@ -478,14 +483,23 @@ def emit(self, invocation: LLMInvocation):
                     GenAI.GEN_AI_USAGE_OUTPUT_TOKENS, completion_tokens
                 )
 
+            message_parts: List[Attributes] = []
             for index, message in enumerate(invocation.messages):
-                content = message.content
-                span.set_attribute(f"gen_ai.prompt.{index}.content", content)
-                span.set_attribute(f"gen_ai.prompt.{index}.role", message.type)
+                message_parts.append(message._to_part_dict())
+
+            if len(message_parts) > 0:
+                span.set_attribute("gen_ai.input.messages", message_parts)
+
+            # for index, message in enumerate(invocation.messages):
+            #     content = message.content
+            #     # Set these attributes to upcoming semconv: https://github.com/open-telemetry/semantic-conventions/pull/2179
+            #     span.set_attribute(f"gen_ai.input.messages.{index}.content", [content._to_part_dict()])
+            #     span.set_attribute(f"gen_ai.input.messages.{index}.role", message.type)
 
             for index, chat_generation in enumerate(
                 invocation.chat_generations
             ):
+                # Set these attributes to upcoming semconv: https://github.com/open-telemetry/semantic-conventions/pull/2179
                 span.set_attribute(
                     f"gen_ai.completion.{index}.content",
                     chat_generation.content,

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

@@ -14,7 +14,7 @@
 
 import time
 from dataclasses import dataclass, field
-from typing import List, Optional
+from typing import Any, Dict, List, Optional
 from uuid import UUID
 
 from .data import ChatGeneration, Message
@@ -29,9 +29,9 @@ class LLMInvocation:
     run_id: UUID
     parent_run_id: Optional[UUID] = None
     start_time: float = field(default_factory=time.time)
-    end_time: float = None
+    end_time: Optional[float] = None
     messages: List[Message] = field(default_factory=list)
     chat_generations: List[ChatGeneration] = field(default_factory=list)
-    attributes: dict = field(default_factory=dict)
+    attributes: Dict[str, Any] = field(default_factory=dict)
     span_id: int = 0
     trace_id: int = 0