Learn how to use Portkey’s Universal API to orchestrate multiple LLMs in a structured debate while tracking performance and evaluating outputs with Arize.

Google Collab Link

​
Overview

This guide demonstrates how to:
  • Use Portkey’s Universal API to seamlessly switch between different LLMs (GPT-4, Claude, Gemini)
  • Implement distributed tracing with Arize and OpenTelemetry
  • Build a multi-agent debate system where LLMs take different roles
  • Export traces and run toxicity evaluations on outputs

​
Prerequisites

Before starting, you’ll need:

​
Installation

Install the required packages: 
pip install portkey-ai openinference-instrumentation-portkey arize-otel arize-phoenix "arize[Tracing]>=7.1.0"

​
Setting Up Tracing

First, configure Arize tracing with Portkey’s instrumentor to capture all LLM calls: 
from arize.otel import register
from openinference.instrumentation.portkey import PortkeyInstrumentor

# Setup OpenTelemetry with Arize
tracer_provider = register(
    space_id = os.getenv("ARIZE_SPACE_ID"),
    api_key = os.getenv("ARIZE_API_KEY"),
    project_name = "portkey-debate",
)

# Enable Portkey instrumentation
PortkeyInstrumentor().instrument(tracer_provider=tracer_provider)

​
Implementing the Multi-LLM Debate

Here’s how to set up different LLMs for different roles using Portkey’s Universal API: 
from portkey_ai import Portkey
import os

# Initialize each LLM client with Portkey
PORTKEY_API_KEY = os.getenv("PORTKEY_API_KEY")

# GPT-4 for "against" arguments
openai = Portkey(
    api_key = PORTKEY_API_KEY,
    virtual_key = os.getenv("OPENAI_VIRTUAL_KEY")
)

# Claude for "pro" arguments
claude = Portkey(
    api_key = PORTKEY_API_KEY,
    virtual_key = os.getenv("CLAUDE_VIRTUAL_KEY")
)

# Gemini for moderation & prompt refinement
gemini = Portkey(
    api_key = PORTKEY_API_KEY,
    virtual_key = os.getenv("GEMINI_VIRTUAL_KEY")
)

​
Debate Round Function

Create a function that orchestrates a single debate round: 
def debate_round(topic: str, debate_prompt: str) -> dict:
    """
    Runs one debate round:
    1. Claude makes the PRO argument
    2. GPT-4 makes the CON argument
    3. Gemini scores both and suggests a refined prompt
    """

    # PRO side (Claude)
    pro_resp = claude.chat.completions.create(
        messages = [{
            "role": "user",
            "content": f"Argue in favor of: {topic}\n\nContext: {debate_prompt}"
        }],
        model = "claude-3-opus-20240229",
        max_tokens = 250
    )
    pro_text = pro_resp["choices"][0]["message"]["content"]

    # CON side (GPT-4)
    con_resp = openai.chat.completions.create(
        model = "gpt-4",
        messages = [{
            "role": "user",
            "content": f"Argue against: {topic}\n\nContext: {debate_prompt}"
        }]
    )
    con_text = con_resp["choices"][0]["message"]["content"]

    # Moderator (Gemini)
    mod_resp = gemini.chat.completions.create(
        model = "gemini-1.5-pro",
        messages = [{
            "role": "user",
            "content": f"""You are a debate moderator. Evaluate these arguments on "{topic}":

            PRO: {pro_text}
            CON: {con_text}

            Suggest an improved debate prompt for more balanced arguments."""
        }]
    )
    new_prompt = mod_resp["choices"][0]["message"]["content"].strip()

    return {"pro": pro_text, "con": con_text, "new_prompt": new_prompt}

​
Running Multiple Rounds

Execute the debate across multiple rounds with progressively refined prompts: 
topic = "Implementing a nationwide four-day workweek"
initial_prompt = "Debate the pros and cons of a four-day workweek."
rounds = 3

prompt = initial_prompt
for i in range(1, rounds + 1):
    result = debate_round(topic, prompt)
    print(f"\n── Round {i} ──")
    print("🔵 PRO:", result["pro"])
    print("\n🔴 CON:", result["con"])
    print("\n🛠️ Suggested New Prompt:", result["new_prompt"])
    prompt = result["new_prompt"]

​
Adding Evaluations

After running the debate, evaluate outputs for toxicity using Arize evals:

​
Export Traces to Dataset

from arize.exporter import ArizeExportClient
from arize.utils.types import Environments
from datetime import datetime

client = ArizeExportClient()

# Export traces from Arize
primary_df = client.export_model_to_df(
    space_id='YOUR_SPACE_ID',
    model_id='portkey-debate',
    environment=Environments.TRACING,
    start_time=datetime.fromisoformat('2025-06-19T07:00:00.000+00:00'),
    end_time=datetime.fromisoformat('2025-06-28T06:59:59.999+00:00')
)

primary_df["input"] = primary_df["attributes.input.value"]
primary_df["output"] = primary_df["attributes.output.value"]

​
Run Toxicity Evaluation

from phoenix.evals import (
    TOXICITY_PROMPT_RAILS_MAP,
    TOXICITY_PROMPT_TEMPLATE,
    OpenAIModel,
    llm_classify,
)

# Configure evaluation model
model = OpenAIModel(
    model_name="gpt-4",
    temperature=0.0,
)

# Run toxicity classification
rails = list(TOXICITY_PROMPT_RAILS_MAP.values())
toxic_classifications = llm_classify(
    dataframe=primary_df,
    template=TOXICITY_PROMPT_TEMPLATE,
    model=model,
    rails=rails,
    provide_explanation=True,
)

​
Send Results Back to Arize

from arize.pandas.logger import Client

arize_client = Client(
    space_id=os.getenv("ARIZE_SPACE_ID"),
    api_key=os.getenv("ARIZE_API_KEY")
)

# Log evaluation results
arize_client.log_evaluations_sync(toxic_classifications, 'portkey-debate')

​
Benefits of This Approach

  1. Unified API: Use the same interface for all LLMs, making it easy to switch providers
  2. Automatic Tracing: All LLM calls are automatically traced without modifying your code
  3. Multi-Agent Orchestration: Different LLMs can play different roles based on their strengths
  4. Comprehensive Observability: Monitor latency, costs, and outputs across all providers
  5. Quality Assurance: Automated evaluations ensure outputs meet safety standards

​
Next Steps

  • Try different LLM combinations for various roles
  • Add more evaluation criteria beyond toxicity
  • Implement fallback strategies using Portkey’s gateway features
  • Set up alerts in Arize for performance degradation