Pydantic AI Slim

1.75.0 · active · verified Wed Apr 01

Pydantic AI Slim is an active Python agent framework that provides a slim package designed to use Pydantic with Large Language Models (LLMs), focusing on reduced dependencies. It enables developers to build type-safe LLM agents and handle structured outputs with minimal boilerplate, leveraging Pydantic's data validation capabilities. The library maintains a rapid release cadence, with version 1.75.0 currently available.

Warnings

gotcha API keys must be provided or set as environment variables (e.g., `GOOGLE_API_KEY`, `OPENAI_API_KEY`, `ANTHROPIC_API_KEY`). Forgetting this will result in a `UserError`.
Fix: Set the appropriate API key as an environment variable before running your application, or pass it directly to the model configuration. Refer to the documentation for model-specific environment variable names.
gotcha When using tools, ensure correct decorator usage with `RunContext`. `@agent.tool_plain` does not support `RunContext` as a parameter. `@agent.tool` requires `RunContext[...]` as its first parameter.
Fix: Use `@agent.tool` if your tool function requires `RunContext` and ensure it is the first parameter. Use `@agent.tool_plain` for simpler tools without context.
breaking Version 1.0.0 of the broader `pydantic-ai` project (which `pydantic-ai-slim` is part of) dropped support for Python 3.9. Additionally, many dataclasses were made to require keyword arguments. Version 1.0.1 made a breaking change where `TenacityTransport` and `AsyncTenacityTransport` now require `pydantic_ai.retries.RetryConfig` instead of raw `tenacity.Retrying` objects.
Fix: Ensure your environment uses Python 3.10 or newer. Update dataclass instantiations to use keyword arguments. Refactor retry logic to use `pydantic_ai.retries.RetryConfig` as specified.
gotcha Mixing `async` and `sync` code, especially in environments like Jupyter Notebooks, can lead to `RuntimeError: This event loop is already running`. While modern Jupyter environments support top-level await, legacy setups or specific integrations might require `nest_asyncio`.
Fix: For synchronous execution in environments with existing event loops, use `agent.run_sync()`. Alternatively, for `async` calls, import and apply `nest_asyncio.apply()` at the beginning of your script.
gotcha Pydantic AI frequently introduces 'beta features' (indicated by a `beta` module) in minor releases. These features' APIs and behaviors may not be stable and are subject to change without backward compatibility guarantees until they are moved out of beta.
Fix: Exercise caution when using beta features. Be prepared for potential API changes in subsequent minor releases. Monitor release notes and documentation for updates on beta feature stability.

Install

pip install pydantic-ai-slim Core installation
pip install "pydantic-ai-slim[google]" Install with Google Gemini support
pip install "pydantic-ai-slim[openai,anthropic]" Install with OpenAI and Anthropic support

Imports

Agent
```
from pydantic_ai import Agent
```
The primary interface for interacting with LLMs and defining agents.
BaseModel
```
from pydantic import BaseModel
```
Used for defining structured outputs and data schemas for agents and tools.
RunContext
```
from pydantic_ai import RunContext
```
Used for dependency injection within agent tools and dynamic instructions.

Quickstart

This quickstart demonstrates how to create a basic Pydantic AI agent using a Google Gemini model. It initializes an `Agent` with a specified model and instructions, then runs a synchronous prompt to get a concise response. Ensure your `GOOGLE_API_KEY` environment variable is set for model access.

import os
from pydantic_ai import Agent

# Ensure GOOGLE_API_KEY is set in your environment
# Example: export GOOGLE_API_KEY="your-api-key-here"
api_key = os.environ.get('GOOGLE_API_KEY')
if not api_key:
    print("Warning: GOOGLE_API_KEY environment variable not set. The example may not run.")

agent = Agent(
    "google-gla:gemini-1.5-flash",
    instructions="You're a helpful assistant. Reply concisely in one sentence."
)

result = agent.run_sync("What is the capital of France?")
print(result.output)
# Expected output: 'The capital of France is Paris.'

view raw JSON →