Phoenix¶

Arize Phoenix is an open-source AI observability platform for evaluating, troubleshooting, and fine-tuning LLM applications. Provides real-time tracing, embedding visualization, and evaluation tools.


Category	observability
Official Docs	Phoenix Documentation
CLI Install	`./cli o11y phoenix install`
CLI Uninstall	`./cli o11y phoenix uninstall`
Namespace	`phoenix`

Overview¶

Phoenix provides AI-native observability: - LLM Tracing: OpenTelemetry-based tracing for LLM calls - Embedding Analysis: Visualize and analyze embedding drift - Evaluations: LLM-as-judge and custom evaluation metrics - Retrieval Analysis: Debug RAG pipeline retrieval quality - Datasets: Create and manage evaluation datasets - Experiments: Run and compare prompt experiments

Installation¶

./cli o11y phoenix install

The installer: 1. Renders Helm values with domain configuration 2. Deploys Phoenix using the official OCI Helm chart 3. Creates the phoenix namespace

No additional environment variables are required.

Verification¶

# Check pods
kubectl get pods -n phoenix

# Check service
kubectl get svc -n phoenix

# Port-forward for UI access
kubectl port-forward svc/phoenix 6006:6006 -n phoenix --address 0.0.0.0 &

# Access Phoenix UI
open http://localhost:6006

Configuration¶

Environment Variables¶

Variable	Description
`DOMAIN`	Domain for ingress (optional)

Usage¶

OpenTelemetry Integration¶

Phoenix supports OpenTelemetry-based instrumentation:

from phoenix.otel import register

tracer_provider = register(
    endpoint="http://phoenix.phoenix:6006/v1/traces"
)

LLM Tracing¶

from openinference.instrumentation.openai import OpenAIInstrumentor

OpenAIInstrumentor().instrument(tracer_provider=tracer_provider)

# All OpenAI calls are now traced
from openai import OpenAI
client = OpenAI(base_url="http://litellm.litellm:4000/v1", api_key="sk-1234")
response = client.chat.completions.create(
    model="vllm/qwen3-30b-instruct-fp8",
    messages=[{"role": "user", "content": "Hello!"}]
)

Embedding Visualization¶

import phoenix as px

px.launch_app()
# Upload embeddings for visualization and drift detection

Troubleshooting¶

UI not loading¶

# Check pod status
kubectl get pods -n phoenix

# Check logs
kubectl logs -n phoenix -l app=phoenix

Traces not appearing¶

# Verify OTLP endpoint is reachable
kubectl exec -it -n <app-namespace> <app-pod> -- \
  curl http://phoenix.phoenix:6006/v1/traces