Thanks, they are coming online as soon as the coding exercise are complete and fully tested. Thanks for the nudge. Happy new 2025 - may this year be the year of learning 😁🎁🎄

I am glad you found the content useful. Happy 2025 and may this year be the year of learning 😁🎁🎄

What version of gpt4all are you using? It was reported earlier that certain versions after 3.4.2 had issues with some of the models. In my case, no special configuration was required.

@@AISoftwareDevelopers I updated to their latest version (1.6.1) It's on windows, I can get my visual models to hallucinate an answer(Based on whatever the address says) but other than that no dice.

I am glad you find it exciting, too! It’s basically one of the most exciting frameworks to come out in 2024. it’s so cool to see a trend towards simplicity. In my opinion, frameworks will shift from DSLs to idiomatic language constructs. And Pydantic is leading the way. Thanks for chiming in 😁👍🎄🎁

@@AngusLou Great question. While not an expert, I can share from experience that CrewAI is more established and it uses domain-specific structures to build and configure agents. PydanticAI is newer and it relies on idiomatic Python code, aiming to simplify the agent management process. Perhaps others, with more experience, can also chime in.

You're welcome! Cheers 😁

@@dreamphoenix thanks for the support

Many corporate users use Azure-based OpenAI APIs as a way to control access, costs and privacy. I used it as an example of an OpenAI-compatible model passed as a client to PydanticAI. Any other model will do just as fine. Thank you for your comment! 👍

Roger that

Thanks, I’m glad you enjoyed the content. It is normal for small creators to struggle with visibility when they first start. Over time, and with comments like this, KZbin starts showing it to more people. Thank you so much for your support. ❤️

Ha! Exactly as you said- PydanticAI uses plain Python, making it look like just a wrapper. Stay tuned for future videos where I’ll show how to customise the system prompts, add tools and agent memory. Then, the power of this simple, yet powerful framework will become more visible. Thanks for the comment. Ciao! 😁🎁🎄

Yes, I will keep that in mind for future videos. Thanks for the suggestion!

Thanks for the support. More complicated demos are on the way. As an introduction, I wanted to keep it simple and show how easy it is to create an agent. 🕵️‍♂️

Thanks for the comment! I’ll double-check the code and see if there’s a fix. In the meantime, if you already solved the problem, feel free to submit a Pull Request so others don’t face the same problem. It will be much appreciated ❤️

Thank you for the support. I am glad you found the content helpful. 😄

There will be more videos, including tools. I have all the code examples ready, it’s just a matter of recording, editing and posting. Please stay tuned and thank you for the support!

@@MajesticGeek Roger that! Will do my best to get them out in a relatively accelerated schedule 🤞🎄🎁

Can't help with that, sorry. Tahnk you for the comment.

Thanks for the sub!

Thank you for the feedback. I am glad you enjoyed the video 😀

Thank you for the support. I am glad you found the content helpful. Cheers, mate! 😄

Hey, thanks for the feedback. I will try to stick to this format, if it helps get the points through easier.

Definitely! A good framework deserves more videos 😁

You are welcome. I am glad you found the content helpful.

The models are stored on my local hard drive - 14'' MPB (late 2023) with M3 and 1TB storage. You can store them in an external drive, if space is an issue. The operations will be slower, but it will work. Thanks for the comment!

Yes, it can be fixed, take a look at the pinned comment where you can enhance the system prompt or introduce a counter in the graph state to prevent runaway conversations among agents. If you need more help, jump on to the discord server and DM me on the tutorial help channel. Cheers!

Glad you enjoyed it

already done see my other post

Yes, as @themax2go pointed out, you can configure and expose an API endpoint and have other apps use the models.

@@AISoftwareDevelopers thanks I will surely try

I am not aware of any limits, but parsing a PDF of that size will be a challenge for any application, not just gpt4all. A powerful CPU, tons of RAM and GPU may be able to help. Otherwise, you may want to parse the PDFs into Markdown first, using something like LlamaParse (paid) and then process the MD files in gpt4all. The embeddings will still take time though.

Yes in options activate openai endpoint

I don’t see a reason why not. The models are downloaded to a folder you can configure and therefore load and use from anywhere else you need to. Great question!

I wasn't aware of this, but after checking they have already released 3 minor updates since the video was recorded. A fast-paced team, for sure 😃

@@AISoftwareDevelopers Those minor updates still don't load most HF models out of the box. Your luck may vary. I use 3.4.2

You can use sideloaded models just fine, but it might require tweaking the chat template. The latest version - 3.6.0 - which was just released does have replacements and examples for several well known sideloaded models.

@@adamtreat7582 thanks for chiming in. What is a good link to learn more about this? If there's enough interest, maybe I can throw together a quick tutorial on how to side-load models?

You really can't using their PC models ... they are very stupid.

Can you elaborate on the use case and the tools? Is this with Nomic Atlas or GPT4ALL?

Yes. The finished product can be tested by checking out the code from GitHub, adding your OpenAI keys and running the Streamlit script. Have you checked out the code? Any issues?

Glad it helped! Care to elaborate on what you enjoyed about the video? It helps small creators like myself make better content.

@@AISoftwareDevelopers Really clearly explained and very logical build up from scratch. Great to follow along. Thank you!

@@AISoftwareDevelopers Will clone the repo next week and try it out!

Great, feel free to DM on the Discord server, if you run into any issues.

@Condinginsight, thank you for the comment.

@Jose-d5h4c, you're absolutely right - multi-agent networks can open possibilities for disruptive use cases, like the travel example you provided. A few years ago Andreessen Horowitz published the now classic "Why Software Is Eating the World" a16z.com/why-software-is-eating-the-world. We are entering a phase where AI will be eating the world, and much sooner than people predict. It took computers and software 70 years to start eating the world, but AI is already almost there in less than 10. But that's not the scariest part. When quantum computing goes mainstream, then a tsunami will unleash, changing paradigms about work and society, well beyond the currently prognosticated scenarios. The final blow will come from the developments in nano technology and robotics, where we will reach the point of being able to grow organic matter and shape it into any robotic being available. Quantum chips, running AGI, inside a completely human-looking and acting robots, that's the true disruption. So, what's to do? While this may sound dark, it's not necessarily bad news. We are nearing the end of the knowledge era that started with Descartes and Gutenberg's press. From the 15th century onwards, we've worshipped at the altar of knowledge. As recently as 20-30 years ago knowledge and hoarding knowledge meant big business. Take the MLS real estate listing services in the States. The fact they hold information meant a whole industry was profiting huge sums of money just by locking out everyone else from the valuable knowledge. But now, knowledge is everywhere. And it is almost free. So, it has become a commodity. The king is dead, long live...wait a second, but what's next? We've entered the age of wisdom. Knowledge alone is not enough anymore - the age of AI will bring us unlimited wisdom. Example: while the MLS can tell a buyer the available homes for sale, their physical properties and price, a true multi-agent AGI will combine millions of data points and understand truly if a home purchase, that home purchase, is the right or the best choice for us. Knowledge has transformed into wisdom. The next natural question is then, with all that wisdom coming [almost] for free, what do we do with it? Well, that will be a deeply personal choice. For my part, I choose to remain optimistic and think all of this will benefit the humanity and make us live in a better world. Thanks for the comment and sorry for the long answer. I hope it made a little bit of sense.

Local agents with Ollama is a great idea. The key will be to find a model that supports agent calling with structured outputs and tools. For small machines the 7B models typically struggle with these tasks, but 37B and 80B models do a better job, for which one needs GPUs. Thanks for the valuable suggestion!

Thanks for the comment. You can start here: techcommunity.microsoft.com/blog/educatordeveloperblog/getting-started-with-azure-ai-studio/4095602 You may also consider using an OpenAI model/key, which may be a bit easier than setting up Azure AI Studio. Both will allow selection of a model and performing the RAG operations in the tutorial, for roughly the same price. I hope this helps.

pip install --upgrade --quiet langchain-community gpt4all Open models.py and add these lines: from langchain_community.llms import GPT4All from gpt4all import Embed4All local_path = ( "./models/Meta-Llama-3-8B-Instruct.Q4_0.gguf" # replace with your local file path ) count = 0 class MyCustomHandler(BaseCallbackHandler): def on_llm_new_token(self, token: str, **kwargs) -> None: global count if count < 10: print(f"Token: {token}") count += 1 self.model_gpt4all = GPT4All( model=local_path, callbacks=[MyCustomHandler()], streaming=True) self.model_gpt4all_embeddings = Embed4All() In ingest.py and chat.py use the model_gpt4all_embeddings and model_gpt4all. That should be it!

Check out discord.gg/5RS7uVpN - it may be easier to message longer text on Discord. Thanks for the question! 😀

Hi @Lirim_K, you bring up an excellent point that I should have addressed in the video. At their current state (as of 12/24) I would not trust any AI-generated code to put in production straight up. No matter how advanced the tool features are - Bolt.new, Cursor, Windsurf...or how much the LLMs claim to understand the intent of the user, as of today, I wouldn't put AI-generated code directly into production. What one could hope for is if AI can reduce the burden on developers or testers, say by X%, that might be a win. With that said, in the example the quality score indeed didn't improve between two runs and the developer agent regenerated the code to get to 950. Final point - the "Quality Score" measure in the tutorial is representative measure and somewhat devoid of a comprehensive quality matrix you might find in static or dynamic code analyzers, such as cyclomatic complexity, potential defects, adherence to requirements, etc. It was used as an example to illustrate the conditional edges in LangGraph and how one can control the agent flow. I hope this answers your question. Thanks so much for chiming in. Make sure to check out the Discord sever at discord.gg/5RS7uVpN for more conversations and insights. Cheers, mate.

Right. with_structured_output is a special method and unfortunately it is not universally supported by all LLMs. the 1 and 8B Ollama models will struggle with this, however the 40B+ might be able to give you an answer. You can also try the mistral models, but I doubt the result will be any different. That is one of the reasons I used an OpenAI model for this tutorial. Normally I prefer local models, but to get structured output, you really need a beefed up LLM. Consider joining the Discord server at discord.gg/5RS7uVpN and we might be able to DM for a deeper discussion.

I am glad you enjoyed the video!

@@mentos.0 thank you for the comment. Let me know if you have any suggestions or questions. Chers, mate! 😁👍

Yes, it could be feasible. Mileage will vary based on the complexity of the tasks, language to implement, LLM to use. But the idea is solid: analyse the big task, break it up to smaller tasks, create code, run the code, export a data file, move to the next agent. Between tasks, I would suggest a quality control agent and conditional edges, to ensure each data export step is done per spec. Start with Python, if that's an option, then use with LangGraph's Plan and Execute template, then integrate REPL into the executor agents and finally use a powerful LLM such as Sonnet 3.5, GPT-4o or Qwen. Watch this guy explain it with a simple example: kzbin.info/www/bejne/Y6rFmISBZad4Y6s. More complex examples will require more rigor, but you get the idea. This is the LangGraph code for the graph: github.com/langchain-ai/langgraph/blob/main/docs/docs/tutorials/plan-and-execute/plan-and-execute.ipynb

To add, consider passing structured output from agent calls and starting with a well-designed GraphState model. Study up Pydantic, it will help greatly when passing structured data among agents.

@@AISoftwareDevelopers Thanks again, appreciate your response, I will check out the resources and will se how it goes.