pydantic-ai-agent-creation

创建 PydanticAI Agent

快速开始

from pydantic_ai import Agent

# Minimal agent (text output)
agent = Agent('openai:gpt-4o')
result = agent.run_sync('Hello!')
print(result.output)  # str

模型选择

模型字符串遵循 provider:model-name 格式：

# OpenAI
agent = Agent('openai:gpt-4o')
agent = Agent('openai:gpt-4o-mini')

# Anthropic
agent = Agent('anthropic:claude-sonnet-4-5')
agent = Agent('anthropic:claude-haiku-4-5')

# Google
agent = Agent('google-gla:gemini-2.0-flash')
agent = Agent('google-vertex:gemini-2.0-flash')

# Others: groq:, mistral:, cohere:, bedrock:, etc.

结构化输出

使用 Pydantic 模型获取经过验证且类型明确的响应：

from pydantic import BaseModel
from pydantic_ai import Agent

class CityInfo(BaseModel):
    city: str
    country: str
    population: int

agent = Agent('openai:gpt-4o', output_type=CityInfo)
result = agent.run_sync('Tell me about Paris')
print(result.output.city)  # "Paris"
print(result.output.population)  # int, validated

Agent 配置

from pydantic_ai import Agent
from pydantic_ai.settings import ModelSettings

agent = Agent(
    'openai:gpt-4o',
    output_type=MyOutput,           # Structured output type
    deps_type=MyDeps,               # Dependency injection type
    instructions='You are helpful.',  # Static instructions
    retries=2,                      # Retry attempts for validation
    name='my-agent',                # For logging/tracing
    model_settings=ModelSettings(   # Provider settings
        temperature=0.7,
        max_tokens=1000
    ),
    end_strategy='early',           # How to handle tool calls with results
)

运行 Agent

三种执行方式：

# Async (preferred)
result = await agent.run('prompt', deps=my_deps)

# Sync (convenience)
result = agent.run_sync('prompt', deps=my_deps)

# Streaming
async with agent.run_stream('prompt') as response:
    async for chunk in response.stream_output():
        print(chunk, end='')

指令与系统提示词

# Instructions: Concatenated, for agent behavior
agent = Agent(
    'openai:gpt-4o',
    instructions='You are a helpful assistant. Be concise.'
)

# Dynamic instructions via decorator
@agent.instructions
def add_context(ctx: RunContext[MyDeps]) -> str:
    return f"User ID: {ctx.deps.user_id}"

# System prompts: Static, for model context
agent = Agent(
    'openai:gpt-4o',
    system_prompt=['You are an expert.', 'Always cite sources.']
)

常见模式

参数化 Agent（类型安全）

from dataclasses import dataclass
from pydantic_ai import Agent, RunContext

@dataclass
class Deps:
    api_key: str
    user_id: int

agent: Agent[Deps, str] = Agent(
    'openai:gpt-4o',
    deps_type=Deps,
)

# deps is now required and type-checked
result = agent.run_sync('Hello', deps=Deps(api_key='...', user_id=123))

无依赖项（满足类型检查器要求）

# Option 1: Explicit type annotation
agent: Agent[None, str] = Agent('openai:gpt-4o')

# Option 2: Pass deps=None
result = agent.run_sync('Hello', deps=None)

验证关卡

在生产代码中使用 Agent 前，请按顺序执行以下步骤：

1. 冒烟测试 — 执行 agent.run_sync('Reply with OK.')（在异步代码中执行 await agent.run(...)）。**通过条件：**调用顺利完成且存在 result.output。
2. 结构化输出 — 如果设置了 output_type，请提示模型生成符合架构的响应。通过条件：result.output 是你的 Pydantic 模型的实例；如果反复验证失败，应收紧指令或调整 retries，而不是立即添加功能。
3. 依赖项 — 如果设置了 deps_type，请使用该类型的 deps= 调用 run / run_sync。**通过条件：**调用通过类型检查并顺利完成（或仅因模型/API 原因失败，而不是因为缺少或错误的 deps 值）。

决策框架

场景	配置
简单文本响应	Agent(model)
结构化数据提取	Agent(model, output_type=MyModel)
需要外部服务	添加 deps_type=MyDeps
需要验证重试	增加 retries=3
调试/监控	设置 instrument=True