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!!
Bro, I have your course and I must say it's amazing. Can you add a section to explain a SQL Agent? Honestly I understand you better than the langgraph guide itself. Thank you very much in advance
Hello Eden, Langgraph is a wonderful tool to create workflows. I was trying to work with payment workflows and came across several challenges. I was working on the the verification example and it seemed like it could not handle failure and exit strategy very well. Could you shed some light on it in your upcoming videos. Would love to see an example workflow for making payments for services based on customer needs.
I'm thrilled with your Udemy course-it's truly impressive! We're dedicated to boosting enrollments, cultivating glowing reviews, and maximizing revenue. I'm eager to brainstorm customized strategies to take your course to even greater heights.
Hello, I was looking at your course on Udemy, you mentioned that you will build apps using python and langchain, however, the student needs to have significant experience in python programming and concepts like "classes". I am disappointed that you mention this because, you can build any app with python without the use of classes for mega projects in ML/DL etc.. That discouraged me from getting your course. You can use Langchain without the need to be an expert python programmer.
hello Eden, pls can you make a tutorial for us on how to use Langgraph Cloud from beginning to end, for example, create a simple AI LangGraph agent and deploy it on LangGraph Cloud or you can just put it in your new course. I have already subscribed.
@@EdenMarco your totally right about RAG. While researching I found out that GraphRAG is promising but it is a new concept from a paper this year: “Retrieval-Augmented Generation with Knowledge Graphs for Customer Service Question Answering” . From what I understand it makes a relational graph where all the data is pre-chunked semantically and doesn’t need to be vectorizes since we wouldn’t need to do vector similarity. Results seemed about 20-40% more accurate answers but with a 10x trade-off in costs and speed.
thanks for this! for production graded SaaS what infrastructure would you suggest ? were looking at DataStax <> Amazon , or possible Azure/Google. Keep it up. ps. is your name the same on linkedin? ps. what is your take on RAGGraphs ?
I study the langchain codebase quite a bit to understand the lessons they're learned and how they've solved them. However, I find langchain to be quite wild and unwieldy and find myself opting to use less and less langchain and more my own abstractions. Langgraph _seems_ be, to me, the approach that Langchain could/should have gone with and I'm finding LG not-too-much-framework.
Until a few years ago, the AI Engineer was supposed to actually train a model (and know how to train and evaluate it in a correct manner and put in production + evaluate while the model is running over time). But today both Software Engineers and Data Scientists need to embrace the advent of the pre-trained models and Gen AI (otherwise they will be useless in 5 years and loose their jobs). So i still think that today's Gen AI engineers are just Software Engineers that know how to put all of the AI components together and just use an API call to the trained AI. Likely they do not know 80% of AI literature amd how to train and build a model from scratch. Unfortunately this will be the direction in this field in the near future (until the AI will take over and these jobs will be useless)
As an example of something you DON'T need a framework to do: if you want to use multiple models you can do that using any routing service, such as OpenRouter, Martian, or BrainTrust. Not only do they handle the model abstraction (generally making every model look like GPT), but they also handle the billing so you don't need N accounts to support N models. If you start development with GPT but want to try out Claude, Gemini, Mistral, etc., this is the easiest way to go.
🎯 Key points for quick navigation: 00:13 *📁 The speaker has been working on a public GitHub repository that implements advanced RAG workflows using LangGraph.* 00:40 *💡 The speaker felt that the existing notebook was missing a software engineering perspective on how to structure an advanced LangGraph application and write maintainable code.* 01:07 *🔩 The speaker refactored the notebook to make it more maintainable, splitting it into sub-modules and writing tests for each chain.* 01:47 *📊 The speaker emphasizes the importance of writing unit tests for code.* 02:44 *🚀 The Advanced RAG workflow involves choosing whether to retrieve documents from a vector store or use a web search, grading documents, and generating an answer while checking for hallucinations and relevance.* 04:23 *💡 The implementation is a combination of three papers on Advanced RAG, corrective RAG, adaptive RAG, and self-RAG.* Made with HARPA AI
To summarize: Don’t build your own software because you’re a moron. Just use this super smart framework from these super smart people. Why reinvent the wheel when someone else is literally reinventing the wheel for you? If LangChain doesn’t do what you need it to do, DONT try to develop something custom or test other frameworks. Instead, just add those features to LangChain using their poorly designed api. Concerned about privacy and vulnerabilities? Fear not, LangChain has explicitly labeled the massive amount of components that are dangerous. Also who do you think you are expecting an open source project to care about your safety. The nerve. This was a great anti-LangChain video. I think I’ll continue to use anything else. Maybe I’ll start with something wild like designing multi-modal apps in python and attaching these revolutionary things called databases so I can integrate my own parsed and formatted data. If I get lucky I might even be able to figure out how to host it all on my own secure servers that don’t expose every console log. Although it might feel a little lonely knowing trackers aren’t watching over me. Who knows though, I’m just a fucking idiot. I should just stick with ChatGPT. I’m sure my company won’t mind if I force feed all of our user data and internal ip into a black box owned by Elon clones.
Why‽ Lang Chain isn't needed if you know how to work with templating, JSON, retrieval, and storage. To be clear, I'm not saying don't use LangCHain. I am saying don't confuse opinionated frameworks for what is right for you. If you like the lying chain approach, go with it for those who have different ideas that are not in line with LangChain or strong opinions. Don't use it; roll your own and share with the community.
Langchain is moving at such breakneck speed with complete disregard to backwards compatibility that the code you wrote couple months ago is obsolete and is not working anymore... Yes it saves you time when you do a quick test but for production, especially if you care about the ability to upgrade your libraries, I'd always build from scratch.
gen ai is in too early for frameworks to be opinionated... learn by experimenting with prompts and Python.. don't use black boxes.. if you're a technical developer, these frameworks won't help you anyway I take exception with Llamaindex pdf reader...
thank you!, one idea I saw and think is a good improvement to the architecture is adding a search into a knwoledge graph module, like dbpedia or similar KGdatabase with the posibilty of adding triplets extracted from the RAG documents itself. The result of the semantic and keyword queries to vectorDb and KGDb will enrich the context provided to the LLM
can't you just combine them both to get the best of both worlds? i guess you could also bind the tools when invoking the react prompt, so that the model would call a necessary tool based on the final result decision?
couldn't agree more. I was having issues using frameworks like crewAI to actual do anything slightly useful. Having more control and giving the LLMs more 'binary' choices seems the way to go at the moment.
💯 This is spot on Eden - LLMs need boundaries to thrive! Langchain/Langraph's elegance is giving devs control to leverage the LLM superpowers safely. 2024 is gonna be the year of the *working* agents thanks to this approach! Great stuff as always, Eden! 🙏
Eden, lang graph doesn't have any good checkpoint libraries apart from sqlite for production use cases like you have for langchain. Do you know anything about that?
Great question ... what about some nosql ways like redis etc ...for checkpointing ... also ended up creating my own way of selecting last K messages ... you can't pass the whole conversational history for a thread to the model (i.e implementing react agent with memory)
Can you make a video going through at a high level each branch in order? Also could you cover LangGraph workflows involving tool use / function calling? Thank you!
