PydanticAI - Building a Research Agent

In this video, I look at how to build a research agent using PydanticAI that makes use of external search as a tool .

In this video, I look at Pydantic AI, which is a new agent framework from the creators of the Pydantic validation framework. It can be used for building a variety of LLM apps and AI agents.

For more tutorials on using LLMs and building agents, check out my Patreon

🕵️ Interested in building LLM Agents? Fill out the form below

👨‍💻Github:

⏱️Time Stamps:
00:00 Intro
00:14 Building a Research Agent with Pydantic AI
00:51 Setting up Search Engines Tavily and Duck Duck Go Search
04:25 Setting up the Agent
05:43 Setting up the Search Agent Tool
07:24 Setting up the Dependencies
08:17 Running a Query
13:25 Injecting a Date into System prompt

[nathannowack]

nice video! you might want to explore Annotated types so you don’t have to waste space in your english prompts saying “3-5 items” etc and instead say Annotated[list[str], Field(description=‘at least 3’, max_length=5)] so you can use that in your data classes / models / typed dicts. then reserve your system prompts for high level nuance

[thembamahlangu]

Nice vid. If you like this checkout Controlflow. Similar features. Better options for orchestration. Good docs. Clean code. Easy to dive in and customise.

[simonNAME]

Can’t wait for the rag video! Thanks, great job

[rezakn]

This was marvellous! Thanks a lot Sam! 🔥🤗

[ShlomiSchwartz]

Awesome video! How can you link Pydantic AI with OpenAI's vision features?

[thanhquachable]

thanks for the video, is there a way to control which tools the agent will call in a specific situation?

[ibrahimhalouane]

Great content! thanks. I'm using Phidata for that purpose and more other stuff, it's relatively lightweight/flexible although it's getting bigger recently, pydantic-ai looks a good alternative when it becomes more mature, having that in mind someone needs to choose carefully what to use in order to reduce the cost of refactoring. I wish to see your thoughts about it or make a video if it worth it.

[TheAstralftw]

one question. why someone would use pytdanticAi instead of langchain for example?

[ksprashutv]

Nice example. I was thinking - I don't like passing the date to the model by overriding the system prompt. What if we give it a tool that allows it to call the datetime function (I don't know what's a python equivalent of moment.js) and it could do all sorts of date-time computations depending on what is asked in the user prompt. That way it would truly be agentic.

[xTomster]

Liked and subscribe after discover the 2 latest videos on Pydantic AI

[justindressler]

Researching at 50 giga pages per minute cool. 😊

[ShaneHolloman]

Sam, I love your work! I am hoping your RAG chunking strategy will be full Page or Article-Based chunks?

[fontenbleau]

Researching what is question. Patent can be registered only to genuine author, which Ai cannot be and human using Ai are not author. At the same time they registered a thousand machine generated patents in USA already.

[MariuszWoloszyn]

Hi. The dataclass that you used is _not_ a pydantic dataclass. For pydantic dataclasss you would: from pydantic.dataclasses import dataclass not from dataclasses import dataclass. It's kind of important difference.