Skip to main content
Arize Phoenix provides open-source AI observability with powerful visualization and debugging tools. Combined with Portkey, get automatic trace collection plus gateway features like caching, fallbacks, and load balancing.

​
Why Phoenix + Portkey?

Visual Debugging

Powerful UI for exploring traces, spans, and debugging LLM behavior

OpenInference Standard

Industry-standard semantic conventions for AI/LLM observability

Evaluation Tools

Built-in tools for evaluating model performance and behavior

Gateway Intelligence

Portkey adds caching, fallbacks, and load balancing to every request

​
Quick Start

pip install arize-phoenix-otel openai openinference-instrumentation-openai

import os
from phoenix.otel import register
from openinference.instrumentation.openai import OpenAIInstrumentor
from openai import OpenAI

# Send traces to Portkey
os.environ["OTEL_EXPORTER_OTLP_ENDPOINT"] = "https://api.portkey.ai/v1/logs/otel"
os.environ["OTEL_EXPORTER_OTLP_HEADERS"] = "x-portkey-api-key=YOUR_PORTKEY_API_KEY"

# Initialize Phoenix
register(set_global_tracer_provider=False)
OpenAIInstrumentor().instrument()

# Use Portkey gateway
client = OpenAI(
    api_key="YOUR_PORTKEY_API_KEY",
    base_url="https://api.portkey.ai/v1"
)

response = client.chat.completions.create(
    model="@openai-prod/gpt-4.1",  # Provider slug from Model Catalog
    messages=[{"role": "user", "content": "Hello!"}]
)

print(response.choices[0].message.content)
OpenTelemetry traces in Portkey

​
Setup

  1. Add provider in Model Catalog → get provider slug (e.g., @openai-prod)
  2. Get Portkey API key
  3. Use model="@provider-slug/model-name" in requests

​
What Gets Captured

Phoenix uses OpenInference semantic conventions:
  • Messages: Full conversation history with roles and content
  • Model Info: Model name, temperature, and parameters
  • Token Usage: Input/output token counts for cost tracking
  • Errors: Detailed error information when requests fail
  • Latency: End-to-end request timing
Supported providers: OpenAI, Anthropic, Bedrock, Vertex AI, Azure OpenAI, and more.

​
Configuration Options

​
Custom Span Attributes

Add custom attributes to your traces: 
from opentelemetry import trace

tracer = trace.get_tracer(__name__)

with tracer.start_as_current_span("custom_operation") as span:
    span.set_attribute("user.id", "user123")
    span.set_attribute("session.id", "session456")
    response = client.chat.completions.create(...)

​
Sampling Configuration

Control trace sampling for production: 
from opentelemetry.sdk.trace.sampling import TraceIdRatioBased

# Sample 10% of traces
register(
    set_global_tracer_provider=False,
    sampler=TraceIdRatioBased(0.1)
)

​
Troubleshooting

Ensure both OTEL_EXPORTER_OTLP_ENDPOINT and OTEL_EXPORTER_OTLP_HEADERS are correctly set before initializing Phoenix.
Call OpenAIInstrumentor().instrument() before creating your OpenAI client.

​
Next Steps

Gateway Configs

Fallbacks, caching, and load balancing

Model Catalog

Manage providers and credentials

Analytics

Cost and performance insights

Phoenix Docs

Official documentation

​
See Your Traces in Action

Once configured, view your Phoenix instrumentation combined with Portkey gateway intelligence in the Portkey dashboard:
OpenTelemetry traces in Portkey