PydanticAI Deep Research

A toy project reimplementing open_deep_research with pydantic-ai. All code logic and prompts are from the original project - I'm just playing around and learning

⚠️ This is just for fun and learning! If you need something reliable, please use the original project.

Installation

This project uses uv as the package manager. Simply run:

uv sync

Quick Start

1. You'll need these API keys:

   • Tavily API key (required for web search)
   • Choose at least one of these:
     • OpenAI API key (if using OpenAI as provider)
     • Anthropic API key (if using Anthropic as provider)

2. Copy .env.example to .env, configure your settings:

# Required
TAVILY_API_KEY=your-tavily-key

# Choose your providers and fill corresponding API keys
PLANNER_PROVIDER=openai     # or anthropic
WRITER_PROVIDER=anthropic   # or openai

# If using OpenAI
OPENAI_API_KEY=sk-xxx
OPENAI_BASE_URL=           # optional, default is official API

# If using Anthropic
ANTHROPIC_API_KEY=xxx
ANTHROPIC_BASE_URL=        # optional

# Optional settings
PLANNER_MODEL=o3-mini              # model for planning
WRITER_MODEL=claude-3.5-sonnet     # model for writing
NUMBER_OF_QUERIES=3                # searches per section
MAX_SEARCH_DEPTH=2                 # max research iterations
MAX_RETRIES=3                      # API call retries
REPORT_STRUCTURE=                  # custom report template

3. Run it:

python cli.py "your topic"

Demo & Logs

Here's what happens when you run it:

All operations are logged to logfire, where you can track the execution flow:

The logs show:

• Search queries generation
• Web search operations
• Section writing progress
• Model API calls

Known Issues

Well... pretty much everything 😂

• Quality is totally unpredictable
• Error handling is minimal
• Code is messy
• Test coverage? What's that?

Credits

Huge thanks to open_deep_research! This is just a learning exercise based on their amazing work.

License

MIT (do whatever you want, but don't blame me if it breaks 😉)