[BUG]: LLMObs.with_annotation_context does not work with Langchain structured outputs

[aW3st]

Tracer Version(s)

3.16.0

Python Version(s)

3.11.2

Pip Version(s)

24.0

Bug Report

Using structured outputs with Langchain seems to break LLMObs's with_annotation_context context manager, such that only the first call is tagged. See code below.

Reproduction Code

Outputs get tagged correctly:

from ddtrace.llmobs import LLMObs
from langchain_openai import AzureChatOpenAI


with LLMObs.annotation_context(
    tags = {
        "sample tag": "works!",
    },
):
    gpt_4 = AzureChatOpenAI(
        openai_api_key=...,
        temperature=0.0,
        model_name="gpt-4.1",
        deployment_name=...,
        azure_endpoint=...,
        openai_api_version=...,
        openai_api_type=...,
    )

    outputs = gpt_4.batch(["Hi how are you? say something nice about the weather"]*5)
    outputs2 = gpt_4.batch(["Hi how are you? say something nice about the sun"]*5)

Only first batch call is tagged:

from pydantic import BaseModel, Field

class Schema(BaseModel):
    user_response: str = Field(..., description="Respond to the User nicely")


with LLMObs.annotation_context(
    tags = {
        "sample tag": "doesn't work :(",
    },
):
    gpt_4 = AzureChatOpenAI(
        openai_api_key=...,
        temperature=0.0,
        model_name="gpt-4.1",
        deployment_name=...,
        azure_endpoint=...,
        openai_api_version=...,
        openai_api_type=...,
    ).with_structured_output(Schema)

    outputs = gpt_4.batch(["Hi how are you? say something nice about the weather"]*5)
    outputs2 = gpt_4.batch(["Hi how are you? say something nice about the sun"]*5)

Error Logs

None

Libraries in Use

pydantic == "2.11.9"
langchain_openai == "0.3.8"

Operating System

No response

[aW3st]

Upgraded to 3.17.0 to see if that would fix, same issue.