🐛 Bug Report: span attributes gen_ai.prompt and gen_ai.completion are deprecated in the latest OpenTelemetry Semantic Conventions

[Kotaro7750]

Which component is this bug for?

LLM Semantic Conventions

📜 Description

In the OpenTelemetry instrumentations in this repository, prompt contents are currently recorded under the following span attributes:

• gen_ai.prompt for request prompts
• gen_ai.completion for response outputs

However, in the latest (v1.38.0) OpenTelemetry GenAI Semantic Conventions, both of these attributes are explicitly marked as deprecated and removed from the specification.

👟 Reproduction steps

Here is sample code using langchain and gemini.

Note that langchain and gemini are just example, and same issue can be said in another tools.

from langchain_google_genai import ChatGoogleGenerativeAI
from langchain_core.messages import SystemMessage
from langchain_core.prompts import ChatPromptTemplate, HumanMessagePromptTemplate
from langchain_core.runnables import RunnablePassthrough


from opentelemetry import trace, _logs, metrics
# Trace
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.export import BatchSpanProcessor
from opentelemetry.exporter.otlp.proto.grpc.trace_exporter import OTLPSpanExporter

# LLM
from opentelemetry.instrumentation.langchain import LangchainInstrumentor
from opentelemetry.instrumentation.google_generativeai import GoogleGenerativeAiInstrumentor


def setUpOpenTelemetry():
    otlp_endpoint = "http://otel-collector:4317/v1"

    # Trace
    provider = TracerProvider()
    processor = BatchSpanProcessor(OTLPSpanExporter(
        endpoint=f'{otlp_endpoint}/traces', insecure=True))
    provider.add_span_processor(processor)
    trace.set_tracer_provider(provider)

    LangchainInstrumentor().instrument()
    GoogleGenerativeAiInstrumentor().instrument()


setUpOpenTelemetry()

template = ChatPromptTemplate.from_messages([
    SystemMessage(content="""Answer the question based on the context below. If the question cannot be answered using the information provided, answer with "I don't know"."""),
    HumanMessagePromptTemplate.from_template("Context: {context}"),
    HumanMessagePromptTemplate.from_template("Questtion: {question}"),
])

model = ChatGoogleGenerativeAI(model="gemini-2.0-flash-lite")


ramen_chatbot = (
    {
        "question": RunnablePassthrough(),
        "context": lambda _: "The capital of France, Japan, and Germany are Paris, Tokyo, and Berlin respectively.",
    }
    | template
    | model
)
response = ramen_chatbot.invoke("What is the capital of Japan?")

print(response.text)

👍 Expected behavior

Although the specification refers to these deprecated attributes as "no replacement at this time", I believe the following fields provide a suitable alternative according to the latest GenAI semantic conventions:

• gen_ai.input.messages for representing model input content
• gen_ai.output.messages for representing model outputs
• gen_ai.system_instructions for system- or model-level instructions

These attributes follow the new message-structured model introduced in the GenAI specification and align with the recommended approach for capturing prompt and response data, instead of using the deprecated gen_ai.prompt and gen_ai.completion span attributes.

---

I believe we can replace constants such as GEN_AI_PROMPT with an appropriate alternative like GEN_AI_INPUT_MESSAGES.

openllmetry/packages/opentelemetry-instrumentation-langchain/opentelemetry/instrumentation/langchain/span_utils.py

Lines 168 to 172 in 9d8cef6

	_set_span_attribute(
	span,
	f"{GenAIAttributes.GEN_AI_PROMPT}.{i}.role",
	_message_type_to_role(msg.type),
	)

In addition, we should follow the updated value format specified by the GenAI semantic conventions (e.g., for gen_ai.output.messages). The expected structure looks like:

[
  {
    “role”: “assistant”,
    “parts”: [
      {
        “type”: “text”,
        “content”: “The weather in Paris is currently rainy with a temperature of 57°F."
      }
    ],
    “finish_reason”: “stop”
  }
]

👎 Actual Behavior with Screenshots

Span with deprecated attributes are emitted.

🤖 Python Version

3.14.0

📃 Provide any additional context for the Bug.

These fields were deprecated in Semantic Conventions v1.28.0 and in the following commit:

32b75a8d465ddea8af396666cd4020c15f4859e1

Although that commit introduced several event-based attributes such as gen_ai.user.message, these event attributes were later re-organized into a structured messaging model in another commit, which was included in v1.37.0 of the specification.

3e06ddb5fc940eff38f66766372fc7b458d03906

This newer version consolidates prompt, output, and instruction data into attributes like gen_ai.input.messages, gen_ai.output.messages, and gen_ai.system_instructions, replacing earlier experimental event fields.

---

There are a few existing issues in this repository that discuss OpenTelemetry SemConv compliance, but most of them take a broad or high-level perspective. None of them focus specifically on the deprecation of gen_ai.prompt and gen_ai.completion or the migration toward the new structured message attributes defined in the latest GenAI semantic conventions.

Because those issues do not address this concrete and well-scoped problem, opening this issue is still meaningful and necessary to ensure proper alignment with the current specification.

like

• 🐛 Bug Report: incompatibilities with LLM semantics #1455

👀 Have you spent some time to check if this bug has been raised before?

• I checked and didn't find similar issue

Are you willing to submit PR?

None