Build a Customer Support Bot | LangGraph

[@donb5521]

Very interesting non-trivial use case. Love the retrieval of user data and persisting of state. Use of mermaid to visually confirm the graph definition is extremely helpful.

[@alchemication]

Thanks for the video. Very useful thoughts to consider for scaling up. It would be interesting to see how could we add something like memory, so agents understand the bigger context about what the user has done in the past, to personalise the experience.

[@andrebadini3573]

Thank you for providing such valuable and practical tutorials that offer real-world benefits for both users and businesses.

[@zacboyles1396]

This was a great demonstration. Thanks for putting it together, it was really thorough and well done. Was anyone else happy to see as little as possible about runnables? I could be wrong but think LCEL has been a massive detour that set LangChain way back. With this demo and a few others on LangGraph, I’ve started to get the feeling things are coming back together.

[@diegocalderon3221]

I think you made a great point at 5:51 in that adding tools/skills or more agents or decisions can actually work against your goal. I think of this as “convergence” toward the user objective.

[@chorltondragon]

Great video. In a project I've just completed I did see some of the benefits of a multi-agent design (simpler than this one). I also saw some of the limitations of LLMs if you attempt to put everything in a single prompt. This video presents a much more structured way of looking at the problem. Thank-you :)

[@Pure_Science_and_Technology]

I haven’t used any embedding models in Ollama yet. One of the reasons is the TTL. I did notice in the upgrade that we can set the time to live keeping the model loaded for embeddings. .

[@fraririri]

Hi! Thank you so much, this is an excellent example, and the way you explain things is great! As you said, it would be really nice to have an example using a semantic router with an embedding classifier

[@dklmn7462]

Still can't catch what is the role of class CompleteOrEscalate here? Why we are using it as a tool? By my understanding tool supposed to be some functions , but not data class

[@of7104]

When building tools in LangGraph that make API calls back to a Django website where the user is currently authenticated, what is the best way to manage and pass authentication tokens to these tools?

[@maxlgemeinderat9202]

Can you go more into Detail about the Memory checkpoint? I have difficulties to understand how i can use the chat history e.g. In memory history

[@kenchang3456]

Thank you very much. Hell of a video 🙂

[@emiliakoleva3775]

Great tutorial! I would like to see soon some example in a task oriented dialogue

[@mukilloganathan1442]

Love seeing Will on the channel!

[@andreamontefiori5727]

Thank you, really useful, informative and interesting video. I spent the first 18 minutes sweating with battery level angst 😅

[@ANKURDIVEKAR]

Thanks for an awesome tutorial. The github link to the code is broken though.

[@emko6892]

Impressive🎉 Can groqcloud be used due to it faster response alongside an interactive UI

[@keenanfernandes1130]

Is there a way to make LangGraph session based, I have been able to do this with Agents using RunnableWithMessageHistory, but using the Supervisor and Agent I couldn't figure out a way to implement session based converstations/workflows

[@XShollaj]

Thank you for the excellent tutorials. Some constructive feedback though, would be to show more love to open source models , and integrate them more in your tutorials instead of just using OpenAI, Anthropic or other closed source models. Newer models like Llama 3, Mixtral 8x22b are good enough to incorporate on your examples and videos (but also tools).

[@willfu-hinthorn]

:) working on it!

[@_rd_kocaman]

exactly. llama3 is good enough for 90% of use cases

[@StoryWorld_Quiz]

do you have any advice on using other llm models?

[@lavamonkeymc]

Question: If I have a data preprocessing agent that has access to around 20 preprocessing tools, what is the best way to go about executing them on a pandas data frame? Do I have the data frame in the State and then pass that input in the function? Does the agent need to have access to that data frame or can we abstract that?

[@willfu-hinthorn]

Ya I'd put the dataframe in the state in this case. The agent would probably benefit from seeing the table schema (columns) and maybe an example row or two so it knows what types of values lie within it. Re: tool organization. It's likely your agent will struggle a bit with 20 tools to choose from, I'd work on trying to simplify things as much as possible by reducing the number of choices the LLM has to make

[@VaibhavPatil-rx7pc]

Great 🎉

[@dowhistleapp6597]

What happens to message , will it grow forever?

[@ramzirebai3661]

Thank you very much

[@sakshamdutta6366]

how can i deploy a langraph ?

[@mahoanghai3364]

Great tutorial

[@ersaaatmeh9273]

when I am using llama3 or mistral it doesn't recognize the tools, does anyone try it?

[@kexia7276]

Ty nice video

[@AICode-Labs]

Could you add reflection on LangGraph nodes ?

[@willfu-hinthorn]

kzbin.info/www/bejne/rGbcnnWKjbOkqs0

[@kunalsolanki5868]

Did anyone try this with Llama 3?

[@iukeay]

Yep. You will need to be careful with the context window but there is some great work arounds for it . Also need to customize the system prompt a little bit for some of the workflows

[@umaima629]

Is this code available on git? Pls share link

[@Ctenaphora]

Please charge your computer.

[@darkmatter9583]

always the same with the videos or the audio issues, videos are great but even me i would buy for myself a microphone at amazon for his videos because he is really good and have to keep updating the opensource community or raise a crowdfunding to buy him a better microphone

[@darwingli1772]

I tried the notebook and swapped using the OpenAI instead of Claude. But it enters a continuous loop and not output anything except consuming token. Am I missing something?

[@williamhinthorn1409]

Hm I’ll run on other models - got a trace link you can share?

[@Leboniko]

He/she expects to get some kind of feedback/error to work with and now ask for help. Your comment demoralizes progress and curiosity. It's a bully comment. Get off youtube and go build something.

[@Slimshady68356]

​@@choiswimmer man will is 100 times a engineer you ever will be , this code design is best what I can see

[@willfu-hinthorn]

Looks like some checks I added to handle some Claude API inconsistencies didn't play well with OAI - pushed up a fix to make it bit more agnostic to the model provider

[@mohibahmed5098]

Plug in a charger bro!!

[@byeebyte]

🎯 Key Takeaways for quick navigation: 00:44 *🚧 Improving the User Experience of Customer Support Chatbots* 00:46 *💼 Enhanced Control over the User Experience* Made with HARPA AI

[@AdamHeffronAI]

your battery is giving me anxiety

[@EricK-bh2sk]

I have anxiety to watch the video in full screen

[@gezaroth]

valuable content, but im having an error, when i run the first example conversation it says i dont have a backup.sqlite file, and i cant get it, is there any other url? even if i copy the 1st travel2.sqlite and change the name to travel2.backup.sqlite, its not working :( 😢

[@sharofazizmatov1000]

Hello. First of all thank you for this video. I am trying to follow you but when I run part_1 I am getting an error in checkpoints and I stuck there. Can you help me to understand what is happening File C:\Python311\Lib\site-packages\langgraph\channels\base.py:117, in create_checkpoint(checkpoint, channels) 115 """Create a checkpoint for the given channels.""" 116 ts = datetime.now(timezone.utc).isoformat() --> 117 assert ts > checkpoint["ts"], "Timestamps must be monotonically increasing" 118 values: dict[str, Any] = {} 119 for k, v in channels.items(): AssertionError: Timestamps must be monotonically increasing

[@ersaaatmeh9273]

did you solve it?

[@sharofazizmatov1000]

@@ersaaatmeh9273 No. I couldn't find a solution

[@willfu-hinthorn]

@@sharofazizmatov1000 I think we fixed this in the most recent relase. Tl;dr, windows timestamping precision was insufficient for our checkpointer.