I'm making some updates to the LangGraph WebSearch Agent. Here's a summary: - Changes to the agent graph - Refactor of the agent script turning agents into classes. - Small bug fix where the scraped content isn't currently being passed to the reporter agent. Only the chosen page and a high level description is being passed. - Supporting Groq models (and maybe Claude if I have time). I'll be putting a video out on this tomorrow.
Definitely the most comprehensive explanation of Langgraph. I've used it, even made a video, but I learned a lot from how you broke down the state. Going to implement a DevOps type of workflow with this, see how well it does. Thanks!
You have very good in depth knowledge about langgraph. Thanks a lot for wonderful content. It’s really differentiating with other TH-cam contents about langgraph
Wish I could drop more than one like, this video has so much more depth and actual advice than 99% of the articles you find online (most feel awfully AI generated). Explaining the actual reasons you'd use langgraph compared to other frameworks actually saved me days of trying each framework
Fantastic job explaining State :: Read/Write :: Graph - Most Langgraph tutorials out there skip this crucial step which leaves a major hole in truly grasping this concept. I also appreciate the way you organize your scripts. The tip of having a model that outputs JSON only is helpful, too. Great job!
The best explanation i have ever seen about the Ai Agent or the langgraph. Nothing confusing, you are straight to the point and good comparisons among other framworks. Really Good Work 👍
Thank you so much for taking the time to share your thoughts (during the weekend!) breaking it all down so thoroughly. I've been looking forward to spend some quality time searching for a good starting point on agents until, weeks ago, I found your channel and easily became confortable with your syle. It's sunday morning and, - yesss! Let's see what he has to say. - Got yourself a new subscriber! Keep up the good work!👏👏👏
thanks man for explaining these concepts in a very simple terms. I had kinda hard time understanding the state in langgraph, your video really helped me understanding it. Keep em coming!
That's the spirit! The best explanatory video I have ever seen. Thank you for your work. I will be waiting for videos explaining the comparison with other frameworks, whether Crew or Agency swarm or etc, with different examples.
I know right, then all the main ones with millions of views are crap and videos are unfollowable, gonna stick links everywhere for this to help the guy out.
What a coincidence. I was playing today with communication between 3 “experts” by going back and forth manually. I was wondering if i could have automated the workflow. I came across this video by chance, and I am very pleased the way you presented. Thank you!
How can we build human-in-the-loop workflows where the agent-workers prompt the human for input, such as helping the person expand on an idea or clarify their goals before proceeding to subsequent steps in the graph?
I would definitely like to see more videos on integrating Ollama, Gemini, and Claude into a mixed agent workflow. I think using the big context of Gemini for the reviewer, Claude for the reporter, and GPT-3.5/Llama mix as the researcher/planner could be a fun project.
Can someone explain why in graph.py he is importing state from state.py? That state object is empty. My understanding is that the workflow.compile(input) sets the state object. Im missing a piece somewhere.
Based on the code you showed, wouldn't your implementation of updating the state `state = {**state, "planner_response": ai_msg.content}` always overwrite exsting planner_response state? You defined the state to be dictionary of lists, but I didn't see how the lists gets appended.
First thanks for the great video and visualisation of the current topic…my question is if it’s possible with this scheme to build an agent graph where you have an agent node that decides which of three tools it should use in the next step to call a specific function from this decided tool could be called parallel tool decision making
Hey John, thank you for this awesome video/resource! I have a question for us that work with a dB aswell. How would you structure the file system and how would you implement handling dB requests and writes, for example, if you want to store the planner's agent thoughts in a dB, how would you achieve that with your modular approach? Thank you so much I learned a lot from this video
Thanks for sharing your view on Langgraph. As far as your efforts towards an open source solution, this is very exciting with Ollama. Not sure I have any others I could suggest. With the structure you have established in your custom agent, it seems to me you could have a more dynamic approach via a control sheet along with a Flask framework to establiah a RAD build out of the code. Any thoughts on this DRY methodology since there is a Template layout? Thanks again.
If the answer of the complex query was wrong. how would you make sure its correct? how can we increase the accuracy of the result? (I want people to trust AI responses)
knowing what you know now about Langgraph, does your position change re: your suggestion/preference to build custom workflows from scratch? I recall you mentioning that preference in your CrewAI tutorial & you don't mention that at the end of your video here. Has Langgraph converted you to frameworks?
Wow Amazing work with clear explanation. Thank you. I want to try it with Ollama using LLama3 model. Will appreciate if you could create another video on same request. I don't want to give my savings to OpenAI.
Amazing, @Data-Centric. I appreciate this tutorial a lot. You really break down the code in a way that is easily digestible, and it is simply articulate and beautiful code, IMO. The VLLM integration tutorial is highly desired for lame ducks like moi.
I will also love to see the adaptation to ollama ran locally. Can you do a chat application in langgraph using ollama as well so we can see how memory works. Thanks great video
top notch work this is excellent stuff thanks for the video! I'm about to ditch crewai. I want to build agents though with a prompt template so that I can apply llama3 specific prompt template so that it works properly and then I can use it on my own hardware. I see you make your own template within the prompt itself. If you use a template then you can abstract that part from the text part of the prompt and use the same format throughout the conversation.
Would it be possible to actually intelligently detect which model would be sufficient at answering the query? Would tou have any ideas for handling that logic? Thanks for posting, nice and tidy!
It's possible. I think if you took that approach you would need a model intelligent enough to decide which model would be best for answering the query.
Why did the agent graph get the answer wrong? I thought that the point of these agent workflows is to handle complex questions with a higher degree of accuracy.
Great explanation of LangGraph.. But the coding part is kind of overwhelming.. I wish there's GUI for LangGraph soon, so that, nodes can be defined, and then connected visually..
I don't think the disadvantage of agency swarm is OpenAI Assistants API, the beta thing will go away soon. The Open Assistants API also will develop quickly. However, it seems to me that LangGraph lets us really define how are agentic workflow moves and how we provide prompts and context to agents in the chain. I don't like relying on a "CEO" or manager agent to decide on the agentic workflow, it seems likely to get this incorrect from time to time. I have had trouble getting the agents I wanted to be used to get selected were the controlling agent at the top of the hierarchy has a choices. Agency Swarm has some configuration possible on this, but it seems LangGraph is much more in this area.
Language-based AI agents are transformative. While exploring Slang AI, consider alternative tools that might better suit your specific needs in communication and automation.
So there is TH-camrs with 1.5mil subs that do AI and you are 100% better! I can follow you and you make sense. Can we use AI to make predictions based on csv data?
Good explanation, than you. But information that AutoGen state management is just to share everything in a Group Chat is outdated: AutoGen already has Memory Enabled Agent and Memory Manager Agent that are likely even more advanced and automated than LangGraph state management, even though that feature is not 100% production ready most likely. AutoGen also allows to use sequential and programmatic workflows to manage what parts of message history are shared with what agents. - These agent frameworks progress very fast and it is hard to keep up with it for sure.
@@free_thinker4958 the most this program does is assemble ready-made imports that query a database on the cloud. Show me architecture that you built without using a single third party package, and you can count me impressed.
Twice now my really cool value added comments got deleted that took a lot of time and effort. Done with that now, sees all ypuu want is people to say "Excellent, clear, concise explanation. Thank you.". Which is true i wont like. Ill just do that from now on, no added effort, apologies
I have not deleted any comments. TH-cam has it's own comment moderation in place, your comments must have triggered it. I am open to constructive feedback, but let's keep things respectful. Thanks for watching.
@@Data-Centric channels can select for comments to be removed from particular users. My own comment was deleted and the criticism was light as can be, talking about the number of dependencies.
I don't know what makes the Langraph a impressive framework. Nobody wants to see this big response and it keeps giving the same response until the Recursion limit and then it stops deh.
I'm making some updates to the LangGraph WebSearch Agent. Here's a summary:
- Changes to the agent graph
- Refactor of the agent script turning agents into classes.
- Small bug fix where the scraped content isn't currently being passed to the reporter agent. Only the chosen page and a high level description is being passed.
- Supporting Groq models (and maybe Claude if I have time).
I'll be putting a video out on this tomorrow.
be awesome if you can show how to use any llm local or groq cloud and the bits of code we would need to change, I've got an idea but not 100%
and Google Studio and Gemini 1.5 pro plzzz
Excellent concise speech. No extraneous mouth noises. No annoying music. No filler & stutter words. Great visuals. Thank you. 🙏🏼
I simply love the simplicity of the content flow and the project setup 👍
Definitely the most comprehensive explanation of Langgraph. I've used it, even made a video, but I learned a lot from how you broke down the state. Going to implement a DevOps type of workflow with this, see how well it does. Thanks!
You have very good in depth knowledge about langgraph. Thanks a lot for wonderful content. It’s really differentiating with other TH-cam contents about langgraph
Thank you! 🙏 The way you keep circling back and comparing and ranking each paradigm is super helpful please keep doing that.
Wish I could drop more than one like, this video has so much more depth and actual advice than 99% of the articles you find online (most feel awfully AI generated). Explaining the actual reasons you'd use langgraph compared to other frameworks actually saved me days of trying each framework
I feel this is the best video on LangGraph , thank you for explaining the concepts so patiently , keep up the good work .
Fantastic job explaining State :: Read/Write :: Graph - Most Langgraph tutorials out there skip this crucial step which leaves a major hole in truly grasping this concept. I also appreciate the way you organize your scripts. The tip of having a model that outputs JSON only is helpful, too. Great job!
Exactly! This guy is the GOAT!
Absolutely the best source for practical ai programming on TH-cam. Can’t get enough please keep sharing these amazing videos & thought process 🔥
The best explanation i have ever seen about the Ai Agent or the langgraph. Nothing confusing, you are straight to the point and good comparisons among other framworks. Really Good Work 👍
This was super helpful! I've seen a lot of other videos on this and, you presentation was the best I've seen so far. Thanks
The more of your videos I watch, the more I trust your opinion. Solid, honest, no nonsense reviews, great work!
Thank you for the detailed information with lang graph, all other videos are kind of sparse when talking about it. Yours is the best I've seen!
Thank you so much for taking the time to share your thoughts (during the weekend!) breaking it all down so thoroughly. I've been looking forward to spend some quality time searching for a good starting point on agents until, weeks ago, I found your channel and easily became confortable with your syle. It's sunday morning and, - yesss! Let's see what he has to say. - Got yourself a new subscriber! Keep up the good work!👏👏👏
Thanks, glad it has been useful.
thanks man for explaining these concepts in a very simple terms. I had kinda hard time understanding the state in langgraph, your video really helped me understanding it.
Keep em coming!
That's the spirit! The best explanatory video I have ever seen. Thank you for your work. I will be waiting for videos explaining the comparison with other frameworks, whether Crew or Agency swarm or etc, with different examples.
So many props for the simple concept breakdown. Very well done. Thank you so much for creating this video!
Damn, this is some high quality presentation. Subscribed.
Completely underrated channel.
I know right, then all the main ones with millions of views are crap and videos are unfollowable, gonna stick links everywhere for this to help the guy out.
truly- Probably nered some hashtags
Excellent, clear, concise explanation. Thank you.
What a coincidence. I was playing today with communication between 3 “experts” by going back and forth manually. I was wondering if i could have automated the workflow.
I came across this video by chance, and I am very pleased the way you presented. Thank you!
Very well illustrated! Thanks for keeping it simple and clear.
This content is amazing, thank you so much. Very well explained and having the repo is cherry on top!
Thank you. Certainly explained quite a few things that the online documentation is missing.
Great video! Thanks for the well-structured code and clear explanations
Thanks!
No BS just to the point 🎉
How can we build human-in-the-loop workflows where the agent-workers prompt the human for input, such as helping the person expand on an idea or clarify their goals before proceeding to subsequent steps in the graph?
I would definitely like to see more videos on integrating Ollama, Gemini, and Claude into a mixed agent workflow. I think using the big context of Gemini for the reviewer, Claude for the reporter, and GPT-3.5/Llama mix as the researcher/planner could be a fun project.
A few people have expressed interest in Integrations with Ollama etc. I'll do this next!
Great video, as usual. I am keen to give it a go now.
I have been waiting for you to do this video...watching now!
I can't thank you enough for this video. Really elucidating.
Thank you for this. It really, really helps simplify the strategy going into something ambitious.
This is fantastic, thanks so much for putting this together!
Can someone explain why in graph.py he is importing state from state.py? That state object is empty. My understanding is that the workflow.compile(input) sets the state object. Im missing a piece somewhere.
Based on the code you showed, wouldn't your implementation of updating the state `state = {**state, "planner_response": ai_msg.content}` always overwrite exsting planner_response state? You defined the state to be dictionary of lists, but I didn't see how the lists gets appended.
How are you calling the serper tool ... Anyone pls help me
best explanation I saw about this topic
good structure of code, ive recently started with this kind of programming again. This is very helpful to get up to speed
That's a great video tutorial! Thanks for sharing!
loved the video! Is anyone else having problems with API related issues, perhaps conflicts between langchain and openai. I just can't figure it out.
Thanks for this wonderful content , Well explained and Clear
Wow. Amazing stuff. Been waiting on this!
This session is really amazing.
Excellent illustration and well explained. Very useful GitHub repo. By any chance you would make a video on using temporal AI?.
First thanks for the great video and visualisation of the current topic…my question is if it’s possible with this scheme to build an agent graph where you have an agent node that decides which of three tools it should use in the next step to call a specific function from this decided tool could be called parallel tool decision making
Great Video! Have you made any open source version of this. Or any advice on how to do it with open source LLM or things to Change.
This is very helpful. Thank you!
Hey John, thank you for this awesome video/resource! I have a question for us that work with a dB aswell. How would you structure the file system and how would you implement handling dB requests and writes, for example, if you want to store the planner's agent thoughts in a dB, how would you achieve that with your modular approach? Thank you so much I learned a lot from this video
Thanks a lot! Super helpful and even more clear than Udemy courses I purchased on LangGraph
Brilliant! Would you consider making a video using DSPy?
Great content! Absolutely love it .
I added Ollama and SearXNG support and sent a pull request. I still need to update the readme for this configuration.
Valeu!
Thanks for the support!
Perfect video, right on time. Thank you and Respect for the valuable information
Pure 🔥 as usual!
THANK YOU!
Would I be crazy saying LangGraph is like LangChain, but for chaining together multiple Lang Chain Agents?
Thanks for sharing your view on Langgraph.
As far as your efforts towards an open source solution, this is very exciting with Ollama. Not sure I have any others I could suggest.
With the structure you have established in your custom agent, it seems to me you could have a more dynamic approach via a control sheet along with a Flask framework to establiah a RAD build out of the code. Any thoughts on this DRY methodology since there is a Template layout?
Thanks again.
10 minutes in and you have my subscription! One question: Is it possible to create one agent that retrieves relevant content from a vector database?
keep up the great releases!
Underrated video!
If the answer of the complex query was wrong. how would you make sure its correct? how can we increase the accuracy of the result? (I want people to trust AI responses)
knowing what you know now about Langgraph, does your position change re: your suggestion/preference to build custom workflows from scratch? I recall you mentioning that preference in your CrewAI tutorial & you don't mention that at the end of your video here. Has Langgraph converted you to frameworks?
Wow Amazing work with clear explanation. Thank you. I want to try it with Ollama using LLama3 model. Will appreciate if you could create another video on same request. I don't want to give my savings to OpenAI.
I also vote for easy OLLAMA implementation!
Please!
Amazing, @Data-Centric. I appreciate this tutorial a lot. You really break down the code in a way that is easily digestible, and it is simply articulate and beautiful code, IMO. The VLLM integration tutorial is highly desired for lame ducks like moi.
Love your videos! Have you done one on GPT Pilot and Open Devin? If not, please do!
I will also love to see the adaptation to ollama ran locally. Can you do a chat application in langgraph using ollama as well so we can see how memory works. Thanks great video
Thank you for doing these amazing videos.
top notch work this is excellent stuff thanks for the video! I'm about to ditch crewai. I want to build agents though with a prompt template so that I can apply llama3 specific prompt template so that it works properly and then I can use it on my own hardware. I see you make your own template within the prompt itself. If you use a template then you can abstract that part from the text part of the prompt and use the same format throughout the conversation.
Would it be possible to actually intelligently detect which model would be sufficient at answering the query? Would tou have any ideas for handling that logic? Thanks for posting, nice and tidy!
It's possible. I think if you took that approach you would need a model intelligent enough to decide which model would be best for answering the query.
Amazing video!!! subscribed :)
Are u going to do some RAG related video?
Why do use the serper api? Isn't that paid? Why not just plug into regular google?
thanks, this is so clear
Why did the agent graph get the answer wrong? I thought that the point of these agent workflows is to handle complex questions with a higher degree of accuracy.
Great video!
Great explanation of LangGraph.. But the coding part is kind of overwhelming.. I wish there's GUI for LangGraph soon, so that, nodes can be defined, and then connected visually..
I don't think the disadvantage of agency swarm is OpenAI Assistants API, the beta thing will go away soon. The Open Assistants API also will develop quickly. However, it seems to me that LangGraph lets us really define how are agentic workflow moves and how we provide prompts and context to agents in the chain. I don't like relying on a "CEO" or manager agent to decide on the agentic workflow, it seems likely to get this incorrect from time to time. I have had trouble getting the agents I wanted to be used to get selected were the controlling agent at the top of the hierarchy has a choices. Agency Swarm has some configuration possible on this, but it seems LangGraph is much more in this area.
Language-based AI agents are transformative. While exploring Slang AI, consider alternative tools that might better suit your specific needs in communication and automation.
Thank you!
Thank you for your video.
The best concept of LangGraph is the state, very useful. Unfortunately its is very un-pythonic, often times it feels like I’m programming in Json.
hey - good to see you getting some views now. about time. Peace brother
im pretty sure he's an ai
You caught me!
Exactly
Already few reasons to say so
Ollama please, thanks.
finally, someone who doesn't talk too fast.
So there is TH-camrs with 1.5mil subs that do AI and you are 100% better! I can follow you and you make sense. Can we use AI to make predictions based on csv data?
Perhaps you could try Ali babas Qwen2-72B?
Good explanation, than you. But information that AutoGen state management is just to share everything in a Group Chat is outdated: AutoGen already has Memory Enabled Agent and Memory Manager Agent that are likely even more advanced and automated than LangGraph state management, even though that feature is not 100% production ready most likely. AutoGen also allows to use sequential and programmatic workflows to manage what parts of message history are shared with what agents. - These agent frameworks progress very fast and it is hard to keep up with it for sure.
take a look at Buffer of Thoughts framework
It's on my list to research, thanks for the recommendation!
@@Data-Centric are you open for online call consultation?
great explanation but langgraph is complex, hope the library get easier so more developers can play with it.
Please make in it langraph.js
number of dependencies is out of this world
What you mean??
@@free_thinker4958 the most this program does is assemble ready-made imports that query a database on the cloud.
Show me architecture that you built without using a single third party package, and you can count me impressed.
❤
wow
Lang graph vs custom AI
Twice now my really cool value added comments got deleted that took a lot of time and effort. Done with that now, sees all ypuu want is people to say "Excellent, clear, concise explanation. Thank you.". Which is true i wont like. Ill just do that from now on, no added effort, apologies
I have not deleted any comments. TH-cam has it's own comment moderation in place, your comments must have triggered it. I am open to constructive feedback, but let's keep things respectful. Thanks for watching.
@@Data-Centric channels can select for comments to be removed from particular users. My own comment was deleted and the criticism was light as can be, talking about the number of dependencies.
I don't know what makes the Langraph a impressive framework. Nobody wants to see this big response and it keeps giving the same response until the Recursion limit and then it stops deh.
hmmm so why not just use functions in python why langGraph , anyway very clear presentation thanx