OpenTelemetry Vertex AI Instrumentation

Warnings

• breaking The library frequently updates to conform with evolving OpenTelemetry Generative AI Semantic Conventions (GenAI SemConv). This may lead to changes in span attribute names and data structures across minor versions, requiring adjustments in how you query or analyze your traces.
  Fix: Review the changelog and OpenTelemetry GenAI Semantic Conventions documentation for each upgrade to understand changes in attribute naming and structure. Adapt your observability queries or dashboards accordingly.
• gotcha By default, this instrumentation logs prompts, completions, and embeddings to span attributes, which may contain sensitive user data. For privacy reasons or to reduce trace size, you might want to disable this behavior.
  Fix: Set the environment variable `TRACELOOP_TRACE_CONTENT=false` before running your application to prevent the logging of sensitive content to span attributes.
• gotcha This instrumentation relies on the `google-cloud-aiplatform` and `google-generativeai` libraries. Ensure that compatible versions of these libraries are installed alongside the instrumentation to avoid runtime errors or incomplete tracing. Specific compatibility requirements may evolve.
  Fix: Refer to the `pyproject.toml` or `setup.py` of the `opentelemetry-instrumentation-vertexai` project on GitHub for precise version constraints. A common requirement is `google-cloud-aiplatform >= 1.38`.

Install

• pip install opentelemetry-instrumentation-vertexai Install library
• pip install opentelemetry-sdk google-cloud-aiplatform google-generativeai Install OpenTelemetry SDK and Vertex AI dependencies

Imports

• VertexAIInstrumentor

  from opentelemetry.instrumentation.vertexai import VertexAIInstrumentor

  Import the instrumentor class to enable tracing for Vertex AI.

Quickstart

This quickstart demonstrates how to set up a basic OpenTelemetry SDK with a console exporter and enable the `opentelemetry-instrumentation-vertexai` library. It then makes a call to a Vertex AI GenerativeModel. Traces capturing the LLM interaction will be printed to the console. Ensure your Google Cloud authentication is configured (e.g., via `gcloud auth application-default login` or `GOOGLE_APPLICATION_CREDENTIALS`).

import os
import vertexai
from vertexai.generative_models import GenerativeModel
from opentelemetry import trace
from opentelemetry.sdk.resources import Resource
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.export import ConsoleSpanExporter, SimpleSpanProcessor
from opentelemetry.instrumentation.vertexai import VertexAIInstrumentor

# Configure OpenTelemetry SDK
resource = Resource.create({"service.name": "vertexai-app-example"})
tracer_provider = TracerProvider(resource=resource)
tracer_provider.add_span_processor(SimpleSpanProcessor(ConsoleSpanExporter()))
trace.set_tracer_provider(tracer_provider)

# Instrument Vertex AI
VertexAIInstrumentor().instrument()

# Initialize Vertex AI (ensure GOOGLE_APPLICATION_CREDENTIALS or gcloud auth is set)
# For a real application, consider explicit project/location via env vars or arguments
project_id = os.environ.get('GCP_PROJECT_ID', 'your-gcp-project-id')
location = os.environ.get('GCP_LOCATION', 'us-central1')
vertexai.init(project=project_id, location=location)

# Use Vertex AI GenerativeModel
model = GenerativeModel("gemini-1.5-flash")

print(f"\n--- Invoking Vertex AI model ({model.model_name}) ---\n")
response = model.generate_content("Explain the importance of OpenTelemetry in AI observability.")

print("Response from model:")
for part in response.candidates[0].content.parts:
    print(part.text)
print("\n--- Traces should be visible in console output ---\n")