FYI, I'm aware that Jar3d made a mistake by returning the wrong kind of steak in the example. The main point was to demonstrate observability; I have since optimised the prompt to address the issue. Test it out for yourself and let's hear how it's working. With the new LangSmith integration, it should be much easier to adapt and contribute to the project.
bro if it works with open source it works with everything ! because its an agent its doing small tasks all ahciveable : so even a small model can perform ! if the task was only given to a single model it would fall down ! ( hence the same model can be used as a agent ! Seperatig the concerns ( ie each agent has its own clean chat history and nothing to do with any other step) ... If the same model performs every step : the chat history aloe can take it off target !
Some stuff that I think would be great if you can cover in your future videos: 1. Persistence in chainlit 2. llm.py: Structured output with instructor library on openai/claude/open-weight models 3. How langsmith is implemented/how debugging is implemented 4. Thoughts on Langgraph studio & Langgraph cloud
@@Data-Centric So I figured out a lot of stuff own my own as I built my code, but I'll leave a separate commend on what I think will be a great subject for the video, regardless of if it's some changes in your system or it's a feature that would be great to have as an llm chatbot
Did you ditch Neo4J because it can only pull from something like an S3 bucket, and this makes it very slow and annoying to use? I'm dead set on using it for my application and I'm looking forward to seeing how you use it.
Many thanks for providing Jar3d. Question - It seems like there's a contention between trying to make the agents 100% JSON-abiding and ease of use by the human user. Requiring /end and /feedback seems very clunky. Are these command ultimately required to get the agents to "play nice" with the structured data? Would be great if we could remove any commands and get the agent workflow to conversationally know when the human was done with their requirements. Thoughts?
I appreciate your continued work on this project! Thank you for keeping it open source. In the video you discussed getting good output from the model as the reason to use frontier models. I have been having some success with Ollama models, mistral-small:22b, using instructor and pydantic. Actually got the idea from one of your TH-cam peers. I have been getting very consistent results even in Ollama. Thank you again for what you do.
What would we need to have it all running in a local machine and not needing any tokens, as response time is not critical for me? Would love for you to have a series on that.
Nicely done! I'll be interested to see the efficiencies for prompts that you build out. Also, I'm curious to see how/if the newly released prompt caching in the API impacts cost.
Such awesome work and great effort! Thank you for sharing all of it! Did I get it right that Jar3d wasn't looking solely for "A-5" text while scraping selling pages to identify if the offer was eligible for output but checked the feedback online on the quality of the product to mark it as A-5 based on the obtained information instead?
To be honest, I didn't pay close attention to what Jar3d was doing since I was more focused on showing off the observability feature. The google shopping search does return a rating, so it wouldn't surprise me if Jar3d was checking that.
I've been looking at contextual RAG. I wanted to have RAG as an offline option so that when I integrate Ollama/vLLM the agent can run completely offline.
Amazing work thanks so much. But i have one issue. In agent_workpad line 48 i got the error from vscode that TypedDict expected a dict or keyword parameter as secod arg. Pls how to splve that
I was thinking about this too - I think in my own setup I'd demote my o1 planning agent to being a member of the agent team - otherwise you are pinging the most expensive model every time you talk to it. Perhaps only invoke it for genuinely complex requests.
Issue with o1model is that it doesn't support structured outputs. You have to tweak the API call to OpenAI to support 01. Rather than do this, I thought it prudent to wait on the release of their more powerful models that will hopefully support structured outputs.
@@Data-Centric , I think that is a good strategy. OpenAI will most likely solve the challenges around agents, so I anticipate that the future will focus on how to effectively use them and provide them with the tools they need. This will be where most of the attention and time is spent in the near future.
I've had this a lot. I just don't have the best acoustics here so have to use software to adjust the sound. Thanks, still a long way to go with Jar3d but I'm happy with the progress.
![[LIVE] : ONE ลุมพินี 89 | คู่เอก "ยอดไอคิว vs คิริลล์"](http://i.ytimg.com/vi/Z3vYkHJ_5Ig/mqdefault.jpg)
FYI, I'm aware that Jar3d made a mistake by returning the wrong kind of steak in the example. The main point was to demonstrate observability; I have since optimised the prompt to address the issue. Test it out for yourself and let's hear how it's working. With the new LangSmith integration, it should be much easier to adapt and contribute to the project.
bro if it works with open source it works with everything !
because its an agent its doing small tasks all ahciveable : so even a small model can perform !
if the task was only given to a single model it would fall down ! ( hence the same model can be used as a agent ! Seperatig the concerns ( ie each agent has its own clean chat history and nothing to do with any other step) ... If the same model performs every step : the chat history aloe can take it off target !
I really appreciate the "clean" way you deliver information. Very metered, clear, focused. Great job on this content!
Your videoes are pure greatness! Such a great walktrough.
You dont seem to cover chat history in your solutions though?
Some stuff that I think would be great if you can cover in your future videos:
1. Persistence in chainlit
2. llm.py: Structured output with instructor library on openai/claude/open-weight models
3. How langsmith is implemented/how debugging is implemented
4. Thoughts on Langgraph studio & Langgraph cloud
Great work! Can you do detailed walkthrough as well? Your agentic system changed so much over the past few months, it's hard to keep up!
Thanks for watching, I'll put something out (time permitting). Is there anything specifically you would like me to cover?
@@Data-Centric So I figured out a lot of stuff own my own as I built my code, but I'll leave a separate commend on what I think will be a great subject for the video, regardless of if it's some changes in your system or it's a feature that would be great to have as an llm chatbot
Did you ditch Neo4J because it can only pull from something like an S3 bucket, and this makes it very slow and annoying to use? I'm dead set on using it for my application and I'm looking forward to seeing how you use it.
Many thanks for providing Jar3d. Question - It seems like there's a contention between trying to make the agents 100% JSON-abiding and ease of use by the human user. Requiring /end and /feedback seems very clunky. Are these command ultimately required to get the agents to "play nice" with the structured data? Would be great if we could remove any commands and get the agent workflow to conversationally know when the human was done with their requirements. Thoughts?
I appreciate your continued work on this project! Thank you for keeping it open source. In the video you discussed getting good output from the model as the reason to use frontier models. I have been having some success with Ollama models, mistral-small:22b, using instructor and pydantic. Actually got the idea from one of your TH-cam peers. I have been getting very consistent results even in Ollama. Thank you again for what you do.
Thanks for watching. I'll be working on the Ollama integration in the coming weeks.
This is pure gold. Thank you!
What would we need to have it all running in a local machine and not needing any tokens, as response time is not critical for me? Would love for you to have a series on that.
I believe GPT now supports memory caching --caching of previously accessed data, which should eliminate redundancy in the web scraping
Nicely done! I'll be interested to see the efficiencies for prompts that you build out. Also, I'm curious to see how/if the newly released prompt caching in the API impacts cost.
Such awesome work and great effort! Thank you for sharing all of it!
Did I get it right that Jar3d wasn't looking solely for "A-5" text while scraping selling pages to identify if the offer was eligible for output but checked the feedback online on the quality of the product to mark it as A-5 based on the obtained information instead?
To be honest, I didn't pay close attention to what Jar3d was doing since I was more focused on showing off the observability feature. The google shopping search does return a rating, so it wouldn't surprise me if Jar3d was checking that.
great, keep it up, this is valuable to me
this will be great. just a bit of change needed for the rag and make it a contextual rag agent and it will be huge upgrade
I've been looking at contextual RAG. I wanted to have RAG as an offline option so that when I integrate Ollama/vLLM the agent can run completely offline.
Amazing work thanks so much. But i have one issue. In agent_workpad line 48 i got the error from vscode that TypedDict expected a dict or keyword parameter as secod arg. Pls how to splve that
Just came to say that we got our first AI agent up and running and had success completing task unsupervised
Can it scrape information from websites with the protection agains bots?
What about the o1 model for the meta agent?
I was thinking about this too - I think in my own setup I'd demote my o1 planning agent to being a member of the agent team - otherwise you are pinging the most expensive model every time you talk to it. Perhaps only invoke it for genuinely complex requests.
Issue with o1model is that it doesn't support structured outputs. You have to tweak the API call to OpenAI to support 01. Rather than do this, I thought it prudent to wait on the release of their more powerful models that will hopefully support structured outputs.
@@Data-Centric , I think that is a good strategy. OpenAI will most likely solve the challenges around agents, so I anticipate that the future will focus on how to effectively use them and provide them with the tools they need. This will be where most of the attention and time is spent in the near future.
thank you
I keep hitting graph recursion limit and hard exit :(
awesome 👍
bro your noise gating and compression made me think you were an AI yourself at first. awesome content though jar3d is badass.
I've had this a lot. I just don't have the best acoustics here so have to use software to adjust the sound. Thanks, still a long way to go with Jar3d but I'm happy with the progress.
Why do you sound similar to Adam the AI voice...?
Just came here to say: whats so hard about python scripts?😅😅😅
I really appreciate both your hard work and you sharing. I'm just starting out, and am very overwhelmed as a newbie
Thanks for watching. Agents are a new thing, we are all still figuring out a lot so take it bit by bit.