AgentWrite with LangGraph

[@autoflujo]

Great video as always! It would be awesome if you can teach us how to deploy LangGraph agents with some cloud service like AWS or digital ocean.

[@EddyLeeKhane]

just swap out the LLM with the API endpoint?

[@jamessebastian8912]

Great share, Sam. I use your channel to keep updated on whats happening in the LLM world.

[@YellowLemon-f1w]

More langgraph and agent content

[@J3R3MI6]

Dude this was amazing! More on this LongWriting topic would be cool

[@bwilliams060]

This is great, thanks Sam. Looking forward to your videos extending on this with search & human in the loop functionality!

[@EarlZMoade]

Great video, I really like how you structured the project as well (going to copy it)

[@samwitteveenai]

Thanks. Please feel free to copy it and extend it as you wish

[@EddyLeeKhane]

Very good explanation and suggestions Brother even for non technical folks, ive just build a slightly simpler version thanks to you (am not a coder, not an "official" AI researcher either, yet i managed thanks to you)

[@enceladus96]

Best AI channel hands down

[@romaindewolff]

Great content, thanks for sharing! 🚀

[@xspydazx]

good work as always and a simple solution ( university get away with murders !) Tasks are tasks ....lol ! But in this case : dissertations and esays etc are often bloated affairs ...SO the boiler plate itself takes up a specific probable 2000 words : this framework is also supply-able... hence instead of supplying a 2000 word boiler plate , we can make a Special STATE ! in which teh STATE is the boiler plate : So for a dissertation , we have the front page the synopsis, the aknowledgements , the contents , the preamble abstract - the literary review, the mtehodology , the findings, the conclusions , the future , the refferences and sources , the apendixes of diagrams and data sources and datasheets etc : by designing a great contents page each section can be farmed out to an agent , we can even sub context each layer to perform is method of hypothosis , premises can conclusions or contrary and alternative solutions and thoughts and rogue methodology ... etc So the state would have major docstrings ... and a mini knowldge graph ( conceptual ) of how the segments conects to the overal space ! ( without passib=ng the whole space ) each previous section must be passed to the other as wellas methodology produced before findings etc ! but all that is your graph ! So if they actually followed the methodology of the creation of the dissertation or reseach paper ... or even a simple essay ... half of the work is only frameworks .... hence people selling dissertation for some subject they never studied as they can see the ( actuality of filling in the form !)... Hence in any formatted document we can produce it with a state and boiler plate yaml! ... and an associated graph ! of the document structure ! so graohs can assist you but your state is your actual worker ! remebering that agents do not need to always see the whole state to continue or conclude !

[@andydataguy]

I bet people would love to see AI Scientist implemented in LangGraph

[@mohmedbadr1947]

I welling to participate in such a project, i would also suggest storm by standford

[@andydataguy]

@@mohmedbadr1947 STORM is also an excellent choice 💯

[@RealEstate3D]

LongCite seems to be interesting. The problem is that especially in RAG users would like to know where a particular part of the answer comes from in order to build up trust in responses of RAG systems.

[@EddyLeeKhane]

wouldnt that be possible with adding metadata?

[@2wwe2eeedits]

Please make a video on how to install Longwriter and AgentWrite I am new to all of this but if you can, can you make it beginner friendly?

[@LawrenceOrsini]

Love to see this deployed in open webui

[@scitechtalktv9742]

How can I influence the number of paragraphs being generated? Or is that totally in the hands of the LLM during the planning phase? Is it a hard requirement that planning phase is totally the responsibility of the LLM (no human influence in that phase at all), or is some human influence in that phase possible? ? Or is this just a question adding the required number of paragraphs in the initial_prompt (in main.py)?

[@Xposterion]

Though you will not be able to control exactly the number of words but only approximate you could append a sentence to the each time you invoke the prompt to each step.

[@scitechtalktv9742]

@@Xposterion My concern was the number of paragraphs that are generated, that I like to be determined by myself. I have tried that, but putting the required number of paragraphs in the initial_prompt and that works very well ! I think the number of words to be generated in each paragraph is determined by the LLM as it is now. Where do these numbers come from? They are often between 300 and 1000. But could I determine that myself too? Maybe as an addition to the initial_prompt?

[@MrAhsan99]

Thanks man for this video

[@ludovicchenneberg6979]

It would be nice you show us how to implement the reflection agents for transition and the researcher. Would it be nice to add a content validator for quality and another for fact checking?

[@mahoanghai3364]

Awesome content

[@_ai_now_]

Great demo)

[@hqcart1]

how is this different by just prompting : "reply with 10k words about subject XYZ in this format, etc..." ?

[@samwitteveenai]

Most models can output that many tokens and even when they can they tend to output only a fraction of that.

[@hqcart1]

@@samwitteveenai Most can't??? i used chatgpt, sonet and llama3, they all can write long article if you prompt it right with the output structure.

[@samwitteveenai]

@@hqcart1 ok whats your prompt for sonnet for a 5k output?

[@kinnaa1234]

awesome, we reverse this process and summarize long documents, right.

[@reza2kn]

Awesome Sam! ❤ May I reach out with a question on designing a distillation pipeline?

[@samwitteveenai]

can reach me on Linkedin

[@DaighoIgho]

Absolutely fantastic video again. Would this framework be suited to replicate a similar project like AI Scientist by Sekana Ai?

[@samwitteveenai]

Yeah I have actually been building versions of that in LangChain and a few other frameworks

[@DaighoIgho]

@@samwitteveenai Sensei, I would appreciate a tutorial on that in whatever framework you suggest would work best

[@bobjones7274]

@@samwitteveenai Information about doing that is something which I need badly.

[@IdPreferNot1]

I see this framework is going help me beat that free flash 1.5 tier like a rented mule. Definitely prefer videos outside of colab and jupyter notebooks

[@ConAim]

Next stress test .... 100K words :)

[@samwitteveenai]

Great idea!! Let me know how it goes.

[@micbab-vg2mu]

Thanks:)