Building STORM from scratch with LangGraph

AI assistants have great potential to automate web research and content generation. A recent work from Shao et al (STORM: Synthesis of Topic Outlines through Retrieval and Multi-perspective Question Asking) offers a new approach for web research to generate Wikipedia-style articles.
It creates wiki editor personas to ask questions about a user-provided topic and experts aided by search to answer questions about a user-provided topic. The two converse, creating an information rich corpus of questions, answer, and references. Then, use this to flesh out a wiki outline in several stages of refinement.
In this video, we show how to implement these ideas using LangGraph from scratch. In it, we create a well structured wiki focused on "Groq and the future of inference" in less than 5 minutes.
Paper:
arxiv.org/abs/...
Notebook:
github.com/lan...

[@andydataguy]

This and the RAPTOR implementation are some of the coolest projects on your channel. Excited to see what else ya'll have been cooking up lately!

[@jessiondiwangan2591]

Another application of STORM will be automating Report generation given a ton of documents for reference. Replacing wikipedia by the reports. This is just an idea. I dont know if its possible.

[@SimonMariusGalyan]

Thank you for the explainer video! Interesting concept which can be adopted for other tasks as well… 😊

[@pascalmauze]

Great job! I had done more or less the same without langgraph. But this makes it much more elegantly.
Just a technical question: in langgraph, as the graph state is a TypeDict schema, it can't carry the llm not the vectorstore, right?... But, in most nodes, we need to create a Runnable using an llm, and ideally a retriever from a vectorstore. Does this mean that the llm and vectorstore should always remain global variables?

[@avinashjain3191]

Pls make a video on
1) write a blog on a topic ,
-- search across web what are the topics related to it..
--search what are the most questions asked by user over search engine realted to blog
-- cover most of targeted questions in my blog
--in my blog, there should be a index, start, main, images, visually charts, graph end point.., and FAQ
Can you take challenge for this, and make video on this..?

[@Tarun_Mamidi]

Great tutorial!
Can you also show how much it cost overall at the end moving forward?
Appreciate all your work!

[@akhilhem1]

I tried this with OpenAI gpt-4 turbo and the overall cost came to about $0.14

[@Tarun_Mamidi]

@@akhilhem1 That's not bad at all. Thanks for sharing!

[@willfu-hinthorn]

You can see all the cost info in LangSmith if you run!@@Tarun_Mamidi

[@GeoffLadwig]

Nice, Thanks

[@jessiondiwangan2591]

The content for each subsections is short. How can we make it longer?

[@tyessenov]

You can try better prompting or add another writer that will expand the topic in details with unlimited depth, virtually

[@jessiondiwangan2591]

@@tyessenov thanks mate

[@playboy71322]

Could this be used to fill out predefined templates as well? I’m assuming yes if you provide the template and ensure each section of the template is filled out appropriately.

[@willfu-hinthorn]

Totally - rather than prompting for a wiki structure, you could define your own template and have the system fill it in

[@jackmartin1146]

Is there a way to deploy this as langserve API?

[@willfu-hinthorn]

Yes! The graph itself should be deployable directly (after running .compile())

[@bahwah]

Thanks a lot! This is really useful. I am getting the following error while running the code interview_graph.astream(initial_state): Error code: 400 - {'error': {'message': "'Dr. Innovation' does not match '^[a-zA-Z0-9_-]{1,64}$. Any idea what could be the possible reason? I had to change the name of "Subject Matter Expert" to Subject_Matter_Expert. But got stuck again with the editors name.

[@digitalisationofutilities4337]

Did you find a solution to this problem ? Getting the same problem tried fixing but still struggling.

[@codewithbod]

same error

[@GeeZoosPL]

Hello, I also had this problem. I solved it by changing the versions of the langchain packages to older ones, i.e.:
langchain-community 0.0.25
langchain-core 0.1.29
langchain-fireworks 0.1.0
langchain-openai 0.0.8
langgraph 0.0.26
langsmith 0.1.11. @@codewithbod

[@akashsri99]

same error

[@Cuzinz]

same error any idea ?

[@ronen.azachi]

i get error on final step it stays None AttributeError Traceback (most recent call last)
in ()
----> 1 final_state = next(iter(final_step.values()))
AttributeError: 'NoneType' object has no attribute 'values'

[@nicola.lepetit]

I have the same error

[@MrDbiswal]

16:22 I have also same issue. final_step is always None because somehow the END is not coming from graph. Are you able to fix it somehow?

[@fengshi9462]

i have the same error, fully one day wasting on it