🛠 Want to get started with freelancing? Let me help: www.datalumina.com/data-freelancer 📚 Learning Data/AI? Join for free: www.skool.com/data-alchemy 🚀 Building AI apps? Check out: launchpad.datalumina.com/ 💼 Need help with a project? Work with me: www.datalumina.com/solutions
Great overview! I’ve been building something from scratch but am considering integrating parts of PydanticAI. In my use case, the code is designed to help users-specifically those who enjoy automating their workflows but aren’t programmers-operate at the simplest level of abstraction, enabling their agents to communicate and collaborate to accomplish tasks. In other words, I’m not directly exposing the Pydantic concept but aiming to simplify it further and create new abstractions that make automation more accessible for my user base
Thank you for the detailed explanation! It would be helpful to see a comparison between tool/function calling and multi-agent approaches, along with guidance on when to use each.
Great breakdown. I'm all for less abstraction. Having been in JS world for years there are so many frameworks and tools coming out just for the sake of it. Focus on the problem, reverse engineer and pick the right tool. In many cases you don't need the chainsaw to top a pencil.
Great video and thanks for sharing. I think having a leaner stack is better because it’s quite easy for a big stack to introduce dependency conflicts. Also Pydantic’s integration with FastAPI is awesome 😂
Hello Dave Ebbelaar, regarding the temperature, you can provide it like so: agent = Agent(model=model,model_settings={'temperature': 0.1}) but it doesn't actually change the behaviour (right now). I think, that needs to be implemented by the Pydantic team 🤔
It depends on the model. Claude prefers XML while OpenAI prefers Markdown or JSON. With small JSON files, it doesn't really matter, but we've found that the model can sometimes miss information with big nested JSON files. As with anything, you can test and compare for your use case to see if you really need the Markdown conversion. I've also found that Markdown is easier to debug when you're looking at it in your observability platform (like Langfuse). It's even more human readable than JSON. Hope that helps.
LangChain is an entire ecosystem. PydanticAI is a really lean framework for solving specific problems around data validation for LLMs. I prefer this leaner, more simple approach.
@@PriyankBolia LangChain’s ecosystem feels overly complex for me. I’ve faced versioning issues and had to dig through multiple abstraction layers to debug unimplemented features, which made troubleshooting a headache. I’d rather build lightweight, purpose-driven components from scratch. Avoiding frameworks helps keep my projects simpler, faster, and free of unnecessary dependencies.
🛠 Want to get started with freelancing? Let me help: www.datalumina.com/data-freelancer
📚 Learning Data/AI? Join for free: www.skool.com/data-alchemy
🚀 Building AI apps? Check out: launchpad.datalumina.com/
💼 Need help with a project? Work with me: www.datalumina.com/solutions
Thanks, Dave. A tutorial for running multiple agents interacting to reach a specific goal would be great!
We need a full tutorial on how to do evals 🙏
Noted!
Love that you chose to do a video on this. I wouldnt bet against Pydantic and see this a an even better version of Swarm.
Great overview! I’ve been building something from scratch but am considering integrating parts of PydanticAI. In my use case, the code is designed to help users-specifically those who enjoy automating their workflows but aren’t programmers-operate at the simplest level of abstraction, enabling their agents to communicate and collaborate to accomplish tasks. In other words, I’m not directly exposing the Pydantic concept but aiming to simplify it further and create new abstractions that make automation more accessible for my user base
Thank you for the detailed explanation! It would be helpful to see a comparison between tool/function calling and multi-agent approaches, along with guidance on when to use each.
Very informative Dave, thanks for all the work. You're the best
The evaluation is perfect. That's the way. Thanks.
Great breakdown. I'm all for less abstraction. Having been in JS world for years there are so many frameworks and tools coming out just for the sake of it. Focus on the problem, reverse engineer and pick the right tool. In many cases you don't need the chainsaw to top a pencil.
Exactly!
Thank you man, this knowledge is really valuable and presented so well.
Great video and thanks for sharing. I think having a leaner stack is better because it’s quite easy for a big stack to introduce dependency conflicts. Also Pydantic’s integration with FastAPI is awesome 😂
Hello Dave Ebbelaar, regarding the temperature, you can provide it like so: agent = Agent(model=model,model_settings={'temperature': 0.1}) but it doesn't actually change the behaviour (right now). I think, that needs to be implemented by the Pydantic team 🤔
What are real life recommendations to deploy those agents? Especially in a serverless aws env
thanks for all your content! it is very informative and helpful
Thanks!
Dave I really want to know your take on phidata ?
Thanks for your review! What would you recommend to use instead of PydanticAI at the moment (until it's matured)? Just using plain API?
I would like to know that as well
I would like to know this as well
Could you do a video on securing vector embeddings in postgres?
Does it integrate with OpenRouter?
Thanks for all the information I appreciate it.
where is the link to the interactive Jupyter env?
Please use bigger fonts like other channels, sometimes I use a laptop to watch, and its hard to read.
Noted!
Exactly...very difficult to view text on the screen
Great video! I like that you are using the interactive execution in vscode/cursor. How do you debug that code? (I didn't figure that out yet)
The interactive mode is great for debugging as well as you can just go line by line and execute your code.
Hoow is this Agentic Framework comapred to phidata Framework???
or AUTOGEN
I thought llms loved json structure. Cool markdown utility function but why needed?
It depends on the model. Claude prefers XML while OpenAI prefers Markdown or JSON. With small JSON files, it doesn't really matter, but we've found that the model can sometimes miss information with big nested JSON files. As with anything, you can test and compare for your use case to see if you really need the Markdown conversion. I've also found that Markdown is easier to debug when you're looking at it in your observability platform (like Langfuse). It's even more human readable than JSON. Hope that helps.
@@daveebbelaar thank you!
great video.
Does PydanticAI support local running LLM’s?
Yes OpenAI compatible endpoint
@ how about OLLAMA?
Please use bigger fonts. Thanks 👍
How does it compare to langchain?
LangChain is an entire ecosystem. PydanticAI is a really lean framework for solving specific problems around data validation for LLMs. I prefer this leaner, more simple approach.
Skip langchain bro. Trust me. You don't need that pain in your life.
@@TheOrionMusicNetwork 😂
@@daveebbelaar Any specific reasons? problems you faced. Trying learning langraph
@@PriyankBolia LangChain’s ecosystem feels overly complex for me. I’ve faced versioning issues and had to dig through multiple abstraction layers to debug unimplemented features, which made troubleshooting a headache. I’d rather build lightweight, purpose-driven components from scratch. Avoiding frameworks helps keep my projects simpler, faster, and free of unnecessary dependencies.
good tutor
THANKS :)
I can always count on you! 💪🏻
PedanticAI?