Introduction

PydanticAI is a Python agent framework designed to make it less painful to build production-grade applications with Generative AI. It brings the same ergonomic design and developer experience to GenAI that FastAPI brought to web development. Portkey enhances PydanticAI with production-readiness features, turning your experimental agents into robust systems by providing:
  • Complete observability of every agent step, tool use, and interaction
  • Built-in reliability with fallbacks, retries, and load balancing
  • Cost tracking and optimization to manage your AI spend
  • Access to 1600+ LLMs through a single integration
  • Guardrails to keep agent behavior safe and compliant
  • OpenTelemetry integration for comprehensive monitoring

PydanticAI Official Documentation

Learn more about PydanticAI’s core concepts and features

Installation & Setup

1

Install the required packages

pip install -U pydantic-ai openai
2

Configure Portkey Client

Since Portkey is OpenAI SDK compatible, you can use the standard OpenAI client with Portkey’s gateway URL:
from openai import AsyncOpenAI
from pydantic_ai import Agent
from pydantic_ai.models.openai import OpenAIModel
from pydantic_ai.providers.openai import OpenAIProvider

# Set up Portkey client using OpenAI SDK
portkey_client = AsyncOpenAI(
    api_key="YOUR_PORTKEY_API_KEY",
    base_url="https://api.portkey.ai/v1"
)
3

Connect to PydanticAI

After setting up your Portkey client, integrate it with PydanticAI:
# Connect Portkey client to a PydanticAI model
agent = Agent(
    model=OpenAIModel(
        model_name="@openai-team-1/gpt-4o",  # Use Portkey's model format
        provider=OpenAIProvider(openai_client=portkey_client),
    ),
    system_prompt="You are a helpful assistant."
)

Basic Agent Implementation

Let’s create a simple structured output agent with PydanticAI and Portkey. This agent will respond to a query about Formula 1 and return structured data:
from openai import AsyncOpenAI
from pydantic import BaseModel, Field
from pydantic_ai import Agent
from pydantic_ai.models.openai import OpenAIModel
from pydantic_ai.providers.openai import OpenAIProvider

# Set up Portkey client
portkey_client = AsyncOpenAI(
    api_key="YOUR_PORTKEY_API_KEY",
    base_url="https://api.portkey.ai/v1"
)

# Define structured output using Pydantic
class F1GrandPrix(BaseModel):
    gp_name: str = Field(description="Grand Prix name, e.g. `Emilia Romagna Grand Prix`")
    year: int = Field(description="The year of the Grand Prix")
    constructor_winner: str = Field(description="The winning constructor of the Grand Prix")
    podium: list[str] = Field(description="Names of the podium drivers (1st, 2nd, 3rd)")

# Create the agent with structured output type
f1_gp_agent = Agent[None, F1GrandPrix](
    model=OpenAIModel(
        model_name="@openai-team-1/gpt-4o",
        provider=OpenAIProvider(openai_client=portkey_client),
    ),
    output_type=F1GrandPrix,
    system_prompt="Assist the user by providing data about the specified Formula 1 Grand Prix"
)

# Run the agent
async def main():
    result = await f1_gp_agent.run("Las Vegas 2023")
    print(result.output)

if __name__ == "__main__":
    import asyncio
    asyncio.run(main())
The output will be a structured F1GrandPrix object with all fields properly typed and validated:
gp_name='Las Vegas Grand Prix'
year=2023
constructor_winner='Red Bull Racing'
podium=['Max Verstappen', 'Charles Leclerc', 'Sergio Pérez']
You can also use the synchronous API if preferred:
result = f1_gp_agent.run_sync("Las Vegas 2023")
print(result.output)

Advanced Features

Using Portkey Headers for Enhanced Features

When you need additional Portkey features like tracing, metadata, or configs, you can add headers to your client:
from openai import AsyncOpenAI
from pydantic_ai import Agent
from pydantic_ai.models.openai import OpenAIModel
from pydantic_ai.providers.openai import OpenAIProvider

# Set up Portkey client with additional features
portkey_client = AsyncOpenAI(
    api_key="YOUR_PORTKEY_API_KEY",
    base_url="https://api.portkey.ai/v1",
    default_headers={
        "x-portkey-trace-id": "f1-data-request",
        "x-portkey-metadata": '{"app_env": "production", "_user": "user_123"}',
        "x-portkey-config": "your-config-id"
    }
)

# Create agent with enhanced Portkey features
agent = Agent(
    model=OpenAIModel(
        model_name="@openai-team-1/gpt-4o",
        provider=OpenAIProvider(openai_client=portkey_client),
    ),
    system_prompt="You are a helpful assistant."
)

Working with Images

PydanticAI supports multimodal inputs including images. Here’s how to use Portkey with a vision model:
from openai import AsyncOpenAI
from pydantic_ai import Agent, ImageUrl
from pydantic_ai.models.openai import OpenAIModel
from pydantic_ai.providers.openai import OpenAIProvider

# Set up Portkey client
portkey_client = AsyncOpenAI(
    api_key="YOUR_PORTKEY_API_KEY",
    base_url="https://api.portkey.ai/v1",
    default_headers={"x-portkey-trace-id": "vision-request"}
)

# Create a vision-capable agent
vision_agent = Agent(
    model=OpenAIModel(
        model_name="@openai-team-1/gpt-4o",  # Vision-capable model
        provider=OpenAIProvider(openai_client=portkey_client),
    ),
    system_prompt="Analyze images and provide detailed descriptions."
)

# Process an image
result = vision_agent.run_sync([
    'What company is this logo from?',
    ImageUrl(url='https://iili.io/3Hs4FMg.png'),
])
print(result.output)
Visit your Portkey dashboard to see detailed logs of this image analysis request, including token usage and costs.

Tools and Tool Calls

PydanticAI provides a powerful tools system that integrates seamlessly with Portkey. Here’s how to create an agent with tools:
import random
from openai import AsyncOpenAI
from pydantic_ai import Agent, RunContext
from pydantic_ai.models.openai import OpenAIModel
from pydantic_ai.providers.openai import OpenAIProvider

# Set up Portkey client
portkey_client = AsyncOpenAI(
    api_key="YOUR_PORTKEY_API_KEY",
    base_url="https://api.portkey.ai/v1",
    default_headers={"x-portkey-trace-id": "dice-game-session"}
)

# Create an agent with dependency injection (player name)
dice_agent = Agent(
    model=OpenAIModel(
        model_name="@openai-team-1/gpt-4o",
        provider=OpenAIProvider(openai_client=portkey_client),
    ),
    deps_type=str,  # Dependency type (player name as string)
    system_prompt=(
        "You're a dice game host. Roll the die and see if it matches "
        "the user's guess. If so, tell them they're a winner. "
        "Use the player's name in your response."
    ),
)

# Define a plain tool (no context needed)
@dice_agent.tool_plain
def roll_die() -> str:
    """Roll a six-sided die and return the result."""
    return str(random.randint(1, 6))

# Define a tool that uses the dependency
@dice_agent.tool
def get_player_name(ctx: RunContext[str]) -> str:
    """Get the player's name."""
    return ctx.deps

# Run the agent
dice_result = dice_agent.run_sync('My guess is 4', deps='Anne')
print(dice_result.output)
Portkey logs each tool call separately, allowing you to analyze the full execution path of your agent, including both LLM calls and tool invocations.

Multi-agent Applications

PydanticAI excels at creating multi-agent systems where agents can call each other. Here’s how to integrate Portkey with a multi-agent setup: This multi-agent system uses three specialized agents: search_agent - Orchestrates the flow and validates flight selections extraction_agent - Extracts structured flight data from raw text seat_preference_agent - Interprets user’s seat preferences With Portkey integration, you get:
  • Unified tracing across all three agents
  • Token and cost tracking for the entire workflow
  • Ability to set usage limits across the entire system
  • Observability of both AI and human interaction points
Here’s a diagram of how these agents interact:
import datetime
from dataclasses import dataclass
from typing import Literal

from pydantic import BaseModel, Field
from rich.prompt import Prompt

from pydantic_ai import Agent, ModelRetry, RunContext
from pydantic_ai.messages import ModelMessage
from pydantic_ai.usage import Usage, UsageLimits
from openai import AsyncOpenAI

# Set up Portkey clients with shared trace ID for connected tracing
portkey_client = AsyncOpenAI(
    api_key="YOUR_PORTKEY_API_KEY",
    base_url="https://api.portkey.ai/v1",
    default_headers={"x-portkey-trace-id": "flight-booking-session"}
)

# Define structured output types
class FlightDetails(BaseModel):
    """Details of the most suitable flight."""
    flight_number: str
    price: int
    origin: str = Field(description='Three-letter airport code')
    destination: str = Field(description='Three-letter airport code')
    date: datetime.date

class NoFlightFound(BaseModel):
    """When no valid flight is found."""

class SeatPreference(BaseModel):
    row: int = Field(ge=1, le=30)
    seat: Literal['A', 'B', 'C', 'D', 'E', 'F']

class Failed(BaseModel):
    """Unable to extract a seat selection."""

# Dependencies for flight search
@dataclass
class Deps:
    web_page_text: str
    req_origin: str
    req_destination: str
    req_date: datetime.date

# This agent is responsible for controlling the flow of the conversation
from pydantic_ai.models.openai import OpenAIModel
from pydantic_ai.providers.openai import OpenAIProvider

search_agent = Agent[Deps, FlightDetails | NoFlightFound](
    model=OpenAIModel(
        model_name="@openai-team-1/gpt-4o",
        provider=OpenAIProvider(openai_client=portkey_client),
    ),
    output_type=FlightDetails | NoFlightFound,  # type: ignore
    retries=4,
    system_prompt=(
        'Your job is to find the cheapest flight for the user on the given date. '
    ),
    instrument=True,  # Enable instrumentation for better tracing
)

# This agent is responsible for extracting flight details from web page text
extraction_agent = Agent(
    model=OpenAIModel(
        model_name="@openai-team-1/gpt-4o",
        provider=OpenAIProvider(openai_client=portkey_client),
    ),
    output_type=list[FlightDetails],
    system_prompt='Extract all the flight details from the given text.',
)

# This agent is responsible for extracting the user's seat selection
seat_preference_agent = Agent[None, SeatPreference | Failed](
    model=OpenAIModel(
        model_name="@openai-team-1/gpt-4o",
        provider=OpenAIProvider(openai_client=portkey_client),
    ),
    output_type=SeatPreference | Failed,  # type: ignore
    system_prompt=(
        "Extract the user's seat preference. "
        'Seats A and F are window seats. '
        'Row 1 is the front row and has extra leg room. '
        'Rows 14, and 20 also have extra leg room. '
    ),
)

@search_agent.tool
async def extract_flights(ctx: RunContext[Deps]) -> list[FlightDetails]:
    """Get details of all flights."""
    # Pass the usage to track nested agent calls
    result = await extraction_agent.run(ctx.deps.web_page_text, usage=ctx.usage)
    return result.output

@search_agent.output_validator
async def validate_output(
    ctx: RunContext[Deps], output: FlightDetails | NoFlightFound
) -> FlightDetails | NoFlightFound:
    """Procedural validation that the flight meets the constraints."""
    if isinstance(output, NoFlightFound):
        return output

    errors: list[str] = []
    if output.origin != ctx.deps.req_origin:
        errors.append(
            f'Flight should have origin {ctx.deps.req_origin}, not {output.origin}'
        )
    if output.destination != ctx.deps.req_destination:
        errors.append(
            f'Flight should have destination {ctx.deps.req_destination}, not {output.destination}'
        )
    if output.date != ctx.deps.req_date:
        errors.append(f'Flight should be on {ctx.deps.req_date}, not {output.date}')

    if errors:
        raise ModelRetry('\n'.join(errors))
    else:
        return output

# Sample flight data (in a real application, this would be from a web scraper)
flights_web_page = """
1. Flight SFO-AK123
- Price: $350
- Origin: San Francisco International Airport (SFO)
- Destination: Ted Stevens Anchorage International Airport (ANC)
- Date: January 10, 2025

2. Flight SFO-AK456
- Price: $370
- Origin: San Francisco International Airport (SFO)
- Destination: Fairbanks International Airport (FAI)
- Date: January 10, 2025

... more flights ...
"""

# Main application flow
async def main():
    # Restrict how many requests this app can make to the LLM
    usage_limits = UsageLimits(request_limit=15)

    deps = Deps(
        web_page_text=flights_web_page,
        req_origin='SFO',
        req_destination='ANC',
        req_date=datetime.date(2025, 1, 10),
    )
    message_history: list[ModelMessage] | None = None
    usage: Usage = Usage()

    # Run the agent until a satisfactory flight is found
    while True:
        result = await search_agent.run(
            f'Find me a flight from {deps.req_origin} to {deps.req_destination} on {deps.req_date}',
            deps=deps,
            usage=usage,
            message_history=message_history,
            usage_limits=usage_limits,
        )
        if isinstance(result.output, NoFlightFound):
            print('No flight found')
            break
        else:
            flight = result.output
            print(f'Flight found: {flight}')
            answer = Prompt.ask(
                'Do you want to buy this flight, or keep searching? (buy/*search)',
                choices=['buy', 'search', ''],
                show_choices=False,
            )
            if answer == 'buy':
                seat = await find_seat(usage, usage_limits)
                await buy_tickets(flight, seat)
                break
            else:
                message_history = result.all_messages(
                    output_tool_return_content='Please suggest another flight'
                )

async def find_seat(usage: Usage, usage_limits: UsageLimits) -> SeatPreference:
    message_history: list[ModelMessage] | None = None
    while True:
        answer = Prompt.ask('What seat would you like?')
        result = await seat_preference_agent.run(
            answer,
            message_history=message_history,
            usage=usage,
            usage_limits=usage_limits,
        )
        if isinstance(result.output, SeatPreference):
            return result.output
        else:
            print('Could not understand seat preference. Please try again.')
            message_history = result.all_messages()

async def buy_tickets(flight_details: FlightDetails, seat: SeatPreference):
    print(f'Purchasing flight {flight_details=!r} {seat=!r}...')
Portkey preserves all the type safety of PydanticAI while adding production monitoring and reliability.

Production Features

1. Enhanced Observability

Portkey provides comprehensive observability for your PydanticAI agents, helping you understand exactly what’s happening during each execution.
Traces provide a hierarchical view of your agent’s execution, showing the sequence of LLM calls, tool invocations, and state transitions.
# Add trace_id to enable hierarchical tracing in Portkey
portkey_client = AsyncOpenAI(
    api_key="YOUR_PORTKEY_API_KEY",
    base_url="https://api.portkey.ai/v1",
    default_headers={"x-portkey-trace-id": "unique-session-id"}
)

2. OpenTelemetry Integration

For comprehensive monitoring and observability, Portkey supports OpenTelemetry integration through various libraries:
import openlit

# Initialize OpenLit with Portkey's OTEL endpoint
openlit.init(
    otlp_endpoint="https://api.portkey.ai/v1/otel",
    otlp_headers={
        "x-portkey-api-key": "YOUR_PORTKEY_API_KEY"
    }
)

# Your PydanticAI agents will now send telemetry data to Portkey
from openai import AsyncOpenAI
from pydantic_ai import Agent
from pydantic_ai.models.openai import OpenAIModel
from pydantic_ai.providers.openai import OpenAIProvider

portkey_client = AsyncOpenAI(
    api_key="YOUR_PORTKEY_API_KEY",
    base_url="https://api.portkey.ai/v1"
)

agent = Agent(
    model=OpenAIModel(
        model_name="@openai-team-1/gpt-4o",
        provider=OpenAIProvider(openai_client=portkey_client),
    ),
    system_prompt="You are a helpful assistant."
)

OpenTelemetry Documentation

Learn more about Portkey’s OpenTelemetry integration and supported libraries

3. Reliability - Keep Your Agents Running Smoothly

When running agents in production, things can go wrong - API rate limits, network issues, or provider outages. Portkey’s reliability features ensure your agents keep running smoothly even when problems occur. It’s simple to enable fallback in your PydanticAI agents by using a Portkey Config:
from openai import AsyncOpenAI

# Create Portkey client with fallback config
portkey_client = AsyncOpenAI(
    api_key="YOUR_PORTKEY_API_KEY",
    base_url="https://api.portkey.ai/v1",
    default_headers={
        "x-portkey-config": "your-fallback-config-id"
    }
)
This configuration will automatically try Claude if the GPT-4o request fails, ensuring your agent can continue operating.

4. Guardrails for Safe Agents

Guardrails ensure your PydanticAI agents operate safely and respond appropriately in all situations. Why Use Guardrails? PydanticAI agents can experience various failure modes:
  • Generating harmful or inappropriate content
  • Leaking sensitive information like PII
  • Hallucinating incorrect information
  • Generating outputs in incorrect formats
While PydanticAI provides type safety for outputs, Portkey’s guardrails add additional protections for both inputs and outputs. Implementing Guardrails
from openai import AsyncOpenAI
from pydantic_ai import Agent
from pydantic_ai.models.openai import OpenAIModel
from pydantic_ai.providers.openai import OpenAIProvider

# Create Portkey client with guardrails
portkey_client = AsyncOpenAI(
    api_key="YOUR_PORTKEY_API_KEY",
    base_url="https://api.portkey.ai/v1",
    default_headers={
        "x-portkey-config": "your-guardrails-config-id"
    }
)

# Create agent with Portkey-enabled client
agent = Agent(
    model=OpenAIModel(
        model_name="@openai-team-1/gpt-4o",
        provider=OpenAIProvider(openai_client=portkey_client),
    ),
    system_prompt="You are a helpful assistant."
)
Portkey’s guardrails can:
  • Detect and redact PII in both inputs and outputs
  • Filter harmful or inappropriate content
  • Validate response formats against schemas
  • Check for hallucinations against ground truth
  • Apply custom business logic and rules

Learn More About Guardrails

Explore Portkey’s guardrail features to enhance agent safety

5. User Tracking with Metadata

Track individual users through your PydanticAI agents using Portkey’s metadata system. What is Metadata in Portkey? Metadata allows you to associate custom data with each request, enabling filtering, segmentation, and analytics. The special _user field is specifically designed for user tracking.
from openai import AsyncOpenAI
from pydantic_ai import Agent
from pydantic_ai.models.openai import OpenAIModel
from pydantic_ai.providers.openai import OpenAIProvider

# Configure client with user tracking
portkey_client = AsyncOpenAI(
    api_key="YOUR_PORTKEY_API_KEY",
    base_url="https://api.portkey.ai/v1",
    default_headers={
        "x-portkey-metadata": '{"_user": "user_123", "user_tier": "premium", "user_company": "Acme Corp", "session_id": "abc-123"}'
    }
)

# Create agent with Portkey client
agent = Agent(
    model=OpenAIModel(
        model_name="@openai-team-1/gpt-4o",
        provider=OpenAIProvider(openai_client=portkey_client),
    ),
    system_prompt="You are a helpful assistant."
)
Filter Analytics by User With metadata in place, you can filter analytics by user and analyze performance metrics on a per-user basis:

Filter analytics by user

This enables:
  • Per-user cost tracking and budgeting
  • Personalized user analytics
  • Team or organization-level metrics
  • Environment-specific monitoring (staging vs. production)

Learn More About Metadata

Explore how to use custom metadata to enhance your analytics

6. Caching for Efficient Agents

Implement caching to make your PydanticAI agents more efficient and cost-effective:
from openai import AsyncOpenAI
from pydantic_ai import Agent
from pydantic_ai.models.openai import OpenAIModel
from pydantic_ai.providers.openai import OpenAIProvider

# Configure Portkey client with simple caching
portkey_client = AsyncOpenAI(
    api_key="YOUR_PORTKEY_API_KEY",
    base_url="https://api.portkey.ai/v1",
    default_headers={
        "x-portkey-config": "your-simple-cache-config-id"
    }
)

# Create agent with cached LLM calls
agent = Agent(
    model=OpenAIModel(
        model_name="@openai-team-1/gpt-4o",
        provider=OpenAIProvider(openai_client=portkey_client),
    ),
    system_prompt="You are a helpful assistant."
)
Simple caching performs exact matches on input prompts, caching identical requests to avoid redundant model executions.

7. Model Interoperability

PydanticAI supports multiple LLM providers, and Portkey extends this capability by providing access to over 200 LLMs through a unified interface. You can easily switch between different models without changing your core agent logic:
from openai import AsyncOpenAI
from pydantic_ai import Agent
from pydantic_ai.models.openai import OpenAIModel
from pydantic_ai.providers.openai import OpenAIProvider

# OpenAI with Portkey
portkey_openai = AsyncOpenAI(
    api_key="YOUR_PORTKEY_API_KEY",
    base_url="https://api.portkey.ai/v1"
)

# Anthropic with Portkey
portkey_anthropic = AsyncOpenAI(
    api_key="YOUR_PORTKEY_API_KEY",
    base_url="https://api.portkey.ai/v1"
)

# Create agents with different models
openai_agent = Agent(
    model=OpenAIModel(
        model_name="@openai-team-1/gpt-4o",
        provider=OpenAIProvider(openai_client=portkey_openai),
    ),
    system_prompt="You are a helpful assistant."
)

anthropic_agent = Agent(
    model=OpenAIModel(
        model_name="@anthropic-team-1/claude-3-5-sonnet-20241022",
        provider=OpenAIProvider(openai_client=portkey_anthropic),
    ),
    system_prompt="You are a helpful assistant."
)

# Choose which agent to use based on your needs
active_agent = openai_agent  # or anthropic_agent

result = active_agent.run_sync("Tell me about quantum computing.")
print(result.output)
Portkey provides access to LLMs from providers including:
  • OpenAI (GPT-4o, GPT-4 Turbo, etc.)
  • Anthropic (Claude 3.5 Sonnet, Claude 3 Opus, etc.)
  • Mistral AI (Mistral Large, Mistral Medium, etc.)
  • Google Vertex AI (Gemini 1.5 Pro, etc.)
  • Cohere (Command, Command-R, etc.)
  • AWS Bedrock (Claude, Titan, etc.)
  • Local/Private Models

Supported Providers

See the full list of LLM providers supported by Portkey

Set Up Enterprise Governance for PydanticAI

Why Enterprise Governance? If you are using PydanticAI inside your organization, you need to consider several governance aspects:
  • Cost Management: Controlling and tracking AI spending across teams
  • Access Control: Managing which teams can use specific models
  • Usage Analytics: Understanding how AI is being used across the organization
  • Security & Compliance: Maintaining enterprise security standards
  • Reliability: Ensuring consistent service across all users
Portkey adds a comprehensive governance layer to address these enterprise needs. Let’s implement these controls step by step.
1

Create API Key with Config

Since Portkey now uses the model format like @team-name/model-name, you can specify your team and model directly without needing virtual keys. Create a Portkey API key with an attached config:
  1. Go to API Keys in Portkey and Create new API key
  2. Optionally attach a config for advanced routing, fallbacks, and reliability features
  3. Generate and save your API key
2

Configure Model Access

Use Portkey’s model naming format to specify which team/provider can access which models:
from openai import AsyncOpenAI
from pydantic_ai import Agent
from pydantic_ai.models.openai import OpenAIModel
from pydantic_ai.providers.openai import OpenAIProvider

# Configure Portkey client with team-specific model access
portkey_client = AsyncOpenAI(
    api_key="YOUR_PORTKEY_API_KEY",
    base_url="https://api.portkey.ai/v1"
)

# Create agent with team-specific model
agent = Agent(
    model=OpenAIModel(
        model_name="@engineering-team/gpt-4o",  # Team-specific model access
        provider=OpenAIProvider(openai_client=portkey_client),
    ),
    system_prompt="You are a helpful assistant."
)
3

Add Enhanced Features with Headers

For additional governance features like tracing, metadata, and configs, add headers as needed:
from openai import AsyncOpenAI
from pydantic_ai import Agent
from pydantic_ai.models.openai import OpenAIModel
from pydantic_ai.providers.openai import OpenAIProvider

# Configure Portkey client with governance features
portkey_client = AsyncOpenAI(
    api_key="YOUR_PORTKEY_API_KEY",
    base_url="https://api.portkey.ai/v1",
    default_headers={
        "x-portkey-trace-id": "engineering-session",
        "x-portkey-metadata": '{"department": "engineering", "environment": "production"}',
        "x-portkey-config": "your-governance-config-id"
    }
)

# Create agent with governance controls
agent = Agent(
    model=OpenAIModel(
        model_name="@engineering-team/gpt-4o",
        provider=OpenAIProvider(openai_client=portkey_client),
    ),
    system_prompt="You are a helpful assistant."
)

Enterprise Features Now Available

Your PydanticAI integration now has:
  • Team-based model access controls
  • Usage tracking & attribution
  • Governance through configs
  • Security guardrails
  • Reliability features

Frequently Asked Questions

Resources