Dynamic AI Agents with LangGraph, Prompt Engineering Enhancements + RAG

Combining prompt-engineering techniques such as chain-of-reasoning and meta-prompting with Retrieval-Augmented Generation (RAG) on the fly has enabled me to develop a powerful agent for long-running, research-intensive tasks. Jar3d has internet access and significantly enhances tasks like creating newsletters, writing literature reviews, planning holidays, and other research-intensive activities. I will demonstrate Jar3d and explain how it operates at a high level. Jar3d is orchestrated with LangGraph.

Chapters
Introduction: 00:00
Jr3d Demo 02:49
Jar3d Architecture: 18:27
Overview of Jar3d code: 23:39
Prompt Engineering: 31:45
Reviewing Jar3d Newsletter: 44:20
Strengths & Weaknesses: 58:43

[EdFife]

I appreciate your approach and your open source of the code. You have inspired me with some of your other tutorials. Thank you! This framework is less chatty than Autogen and CrewAI. I swear Autogen can do 5 iterations complimenting each other and saying thank you for the feedback. This has amazing amounts of potential. My first plan to extend would be to allow the lower tasks be done with a local model and higher level tasks go through a commercial model. Then maybe a GUI.

[DavidSegura]

🎯 Key points for quick navigation:

00:00:12 *🤖 The speaker introduces "Jared, " an AI agent designed for long-term research tasks using meta prompting, agentic RAG, and chain of reasoning.*
00:01:24 *🛠️ Jared's development and logic are detailed, focusing on meta prompting, agentic RAG, and their implementation through Python code.*
00:03:14 *🐳 Setting up Jared involves configuring an ingestor server via Docker and initializing Jared with specific model choices, facilitating long-term research capabilities.*
00:05:46 *📊 Jared integrates meta prompting to refine goals and gather aligned requirements, employing an iterative chain of reasoning approach to enhance task comprehension.*
00:08:41 *📰 Jared facilitates the creation of concise, informative newsletters by refining goals through meta prompting and tailored questioning, ideal for AI enthusiasts and developers.*
00:12:31 *🔄 Jared's meta expert role orchestrates internet research, writing, and planning tasks based on refined requirements, enhancing workflow efficiency.*
00:18:32 *🛠️ The Jared architecture utilizes LangGraph for workflow orchestration, incorporating state management to track interactions and process outputs effectively.*
24:14 *📊 LangGraph allows recording and accessing various states in workflows, facilitating flexible data handling.*
25:45 *🛠️ The tool expert within the system is complex, involving stages like document ingestion and utilizing a modified Tika server for processing.*
26:56 *📑 RAG on the fly involves document embedding and local model ranking to refine research outputs for meta prompting agents.*
29:15 *🔄 Agent graphs define the workflow sequence from Jared through various expert agents, directed by a router based on meta prompt outputs.*
30:42 *🧠 Setting recursion limits enhances the capability of Jared to manage complex, long-running tasks effectively without needing an infinite context window.*
45:55 *📊 Jared's workflow involves iterative retrieval from diverse sources to gather comprehensive information, essential for creating newsletters.*
46:51 *🌐 Jared's approach mimics extensive web research to compile and synthesize content into coherent newsletters.*
48:42 *📈 Llama 3.1 models offer various sizes and performance benchmarks, showing competitive advantages in AI tasks.*
50:05 *💰 Llama 3.1 models significantly reduce costs compared to other models, making them more accessible for developers.*
51:48 *🛠️ Llama stack API development is mentioned, despite some hallucinations in the source material.*
59:28 *🚫 Jared's limitations include potential crashes beyond 128k context and issues with model convergence in less capable versions like Llama 3.1 70B.*
01:02:18 *🌍 Llama 3.1 405B models facilitate complex workflows like Jared's, enhancing possibilities for enterprise-level AI applications.*

Made with HARPA AI

[fredrikhansen]

Wow, such an inspiration you are! Great work, kudos!

[jakeparker]

Thanks for posting this! Feels good to know using AI for finding information is out in the ether, excited to see what the future of open source brings

[durand]

Amazing workflow you've built! So much to learn from and adapt from Jar3d.

[malikrumi]

Posted 17 minutes ago, and I am here among the first as usual. Teach us!

[ChristopherFoster-McBride]

I enjoyed running this, and your reflections at the end

[jordanallen]

Crazy to think that I made a turd comment on the first video that I saw where you criticized CrewAI. Unsubscribed, and then shortly resubscribed after watching quite a few of your other videos. Sure enough, you were right. Now, I look forward to your content more than the numerous other channels I follow. Keep it going. Extremely helpful for a dev learning AI. And sorry for being a turd

[sebbecht]

really impressive! Definitely want a deeper technical dive on the tool expert.

[RazorCXTechnologies]

Just ran the Meta Agent on my RTX4090 with Llama3.1:70b. It worked great using the Serpa tool. Huge thanks for all your effort!

[SixTimesNine]

Started looking at this yesterday. Very impressive! Code is clean, well written. Like the tika server use. Two feedbacks for you right now: (1) I started with the cli and it went off into an infinite error loop (until it hit 40) - the reason was playwright - it needed 'playwright install' ran, then it worked. (2) in the cli logs I saw plenty of 429s - I think you're making too many requests to something. Don't know if you're trapping that or not (I didn't read the logs closely!). Web version - would be good if somewhere on screen it showed '/end to finish giving requirements' as no one reads the manual... Going to continue playing with it today. So far though - great job! One of the best things I've found in a while.

[jeremybristol]

Great content, interesting idea, appreciate the code walkthrough. At 29:23 you enlarged the font- much easier to read along!

[fpsteiner]

Great inspiration, thank you! I'd be particularly interested in using jared to build his own tools after having determined the detailed specs.

[Alisa-ld]

Hi! Your videos are so detailed and useful! Do you consider a sponsorship?

[SolidBuildersInc]

Jarad,

Thank you for being transparent with your open source model. Fortunately, I believe you can run your solution completely free with a end point deployment. I would like to monitize this solution with you exclusively. The cost to end user would be pennies a day instead of dollars per hr. with a MLM multiplier to help everybody minize cost and be scalable. Let me know if this would be of interest to you.

If so, it's just a matter of

Kudos

[aaagaming]

Great content mate! Youve just earned a new subscriber.

[donconkey]

Great video! I appreciate your methodical approach and precision in developing pipelines and strategies to test the capabilities of SOTA LLMs. Your gentle nudge towards Markdown for prompts is valuable. Have you considered handing off the TLDR LLaMA 3.1 goal to Perplexity to compare the results? Thanks to you, I now better understand why expert prompting is essential for success.

[zmjerry]

This is awesome! Please provide some ways that I can sponsor your development of Jered. CrewAI is burning through my wallet with all the hidden prompts and iterative tokens.

[techme]

Great video!! Thank you for taking the time!

My confusion is…How would I create a multi agent graph where the initial agent asks the user a few questions to determine intent -> based on that it determines what agent to send the user to - this 2nd agent has its own LLM prompt logic -> when this 2nd agent requires feedback from the user … does it communicate with the user directly ? Or does the initial agent only communicate with the user

That is where I’m really confused - any guidance would be great!

Thank you again!!

[saro.saribekyan]

Hello friend. I'm excited about the type of work you are commited to.

Would you be interested to team up and start creating something more sophisticated?

I'm a software engineer with experience in complex projects. I'm highly interested in AI automation topic and want to dig deep and get hands on experience with it. I understand that sometimes ideas are bigger than the time needed to develop something, so I believe teaming up would not only result in better ideas but also the capacity to develop the ideas which were not possible earlier.

Looking forward to your response!