When to use Pydantic AI
- You want type-safe agent development
- You prefer Pydantic’s validation and schema generation
- You want a simpler API than LangChain
- You need structured outputs from your agents
Create an agent
rayai create-agent my_agent --framework pydantic
Tool to execute on Ray workers:
from pydantic_ai import Agent, Tool
from rayai.adapters import from_pydantic_tool
# Existing Pydantic AI Tool
my_tool = Tool(some_function, name="my_tool")
# Wrap for Ray execution
ray_tool = from_pydantic_tool(my_tool, num_cpus=2)
# Use directly with Pydantic AI agent
agent = Agent("openai:gpt-4o-mini", tools=[ray_tool])
from pydantic_ai import Agent
from rayai import tool
from rayai.adapters import RayToolWrapper, AgentFramework
@tool(desc="Search the web", num_cpus=1)
def search_web(query: str) -> str:
return f"Results for: {query}"
# Wrap for Pydantic AI
wrapper = RayToolWrapper(framework=AgentFramework.PYDANTIC)
pydantic_tools = wrapper.wrap_tools([search_web])
# Use with Pydantic AI agent
agent = Agent("openai:gpt-4o-mini", tools=pydantic_tools)
RayToolWrapper auto-detects the source framework and converts tools to Pydantic AI-compatible callables:
from langchain_community.tools import DuckDuckGoSearchRun
from pydantic_ai import Agent
from rayai.adapters import RayToolWrapper, AgentFramework
# LangChain tool
langchain_search = DuckDuckGoSearchRun()
# Plain Python function
def calculate(a: int, b: int) -> int:
"""Add two numbers together."""
return a + b
# Convert ANY tools to Pydantic AI format
wrapper = RayToolWrapper(framework=AgentFramework.PYDANTIC)
tools = wrapper.wrap_tools([langchain_search, calculate], num_cpus=1)
# Use with Pydantic AI agent - all tools execute on Ray
agent = Agent("openai:gpt-4o-mini", tools=tools)
The wrapper auto-detects: Ray @tool functions, LangChain BaseTool, Pydantic AI Tool, and plain Python callables. Type annotations and docstrings are preserved for schema generation.
Next steps
