Learn Pydantic AI agents, step by step,

🤖 Pydantic AI Agents Tutorial

A step-by-step guide to building intelligent AI agents using Pydantic AI and local models and (Ollama or any OAI compatible). This tutorial demonstrates how to create structured, vision-capable, and tool-equipped AI agents.

🌟 Features

• 📝 Structured output
• 🔧 Custom tool integration
• 👁️ Using Vision models (llama3.2 vision / minicpm-v)
• 🤝 Simple Multi-agent systems
• 📁 File handling
• 🔄 Dynamic prompting

🚀 Getting Started

1. Clone the repository
2. Create conda evironment

conda create -n pydantic python=3.12 pip -y

3. Install dependencies:

pip install pydantic-ai llm-sandbox

4. Set up Ollama and required models (chosee any of those model)

# Best agents model for local run
ollama pull llama3.1:8b-instruct-q8_0
ollama pull qwen2.5:14b
ollama pull qwen2.5:32b

# vision models
ollama pull llama3.2-vision:latest
ollama pull minicpm-v

(optional) setup vLLM for tool calling

If you prefer faster local inference you can use vLLM, here is 2 examples for preparing vLLM

Running LLama3.1

# create vllm conda environment 
conda create -n vllm python=3.12 pip -y
conda activate vllm
pip install vllm

# run vllm with LLama3.1 8B with flags that enables tool calling
# assuming you downloaded the GPTQ-Q8 version 

vllm serve Meta-Llama-3.1-8B-Instruct-GPTQ-Q_8 \
           --port 5003 \                                
           --enforce-eager \
           --kv-cache-dtype fp8 \
           --enable-chunked-prefill  \
           --enable-auto-tool-choice \
           --tool-call-parser llama3_json \
           --chat-template ./tool_chat_template_llama3.1_json.jinja

you can download the jinja template from vllm github repo vLLM repo vllm/examples/tool_chat_template_llama3.1_json.jinja

Running Qwen2.5

based on offical documentation Qwen2.5 uses hermes tool calling and hermes jinja template Qwen2.5 function_call

# Running Qwen2.5-32B with dual RTX cards. 

vllm serve Qwen2.5-32B-Instruct-AWQ --port 5003 \                                 
           --tensor-parallel-size 2  \
           --enforce-eager \
           --kv-cache-dtype fp8 \
           --enable-chunked-prefill  \
           --enable-auto-tool-choice \
           --tool-call-parser hermes \
           --chat-template ./tool_chat_template_hermes.jinja

Now all you need is to use OpenAPI model and change the base_url

model = OpenAIModel(
    'Qwen2.5-32B-Instruct-AWQ',
    base_url='http://localhost:5003/v1',
    api_key='vllm',
)

📚 Tutorial Structure

The repository contains progressive examples:

Basic Concepts

• 0_hello_world.py - Basic agent setup
• 1_hello_with_OAI_api.py - OpenAI compatible API integration
• 2_simple_structured.py - Structured outputs
• 3_simple_structured_table.py - Table formatting + Structured outputs

Advanced Features

• 4_lets_make_tools.py - Custom tool creation
• 5_mix_tools_with_structured_output.py - Combining tools and structure
• 6_code_as_tool.py - Code execution capabilities
• 7_code_with_added_libs.py - External library integration

Special Features

• 8_lets_make_dynamic_prompt.py - Dynamic prompt generation
• 9_mix_multiple_prompts.py - Multiple prompt handling
• 10_lets_open_files.py - File operations
• 11_hello_vision.py - Basic vision capabilities
• 12_running_2_mixed_agents_with_vision.py - Advanced vision and multi-agent systems

📝 License

Licensed under the Apache License, Version 2.0. See LICENSE file for details.

🤝 Contributing

Contributions welcome! Please feel free to submit a Pull Request.