Storm: AI 기반 위키백과 생성 도구

스탠퍼드 대학의 AI 프로젝트인 Storm은 웹 페이지에서 참조한 정보를 사용하여 다양한 주제에 대한 Wikipedia 문서를 생성합니다.

### 스탠퍼드의 Storm AI: 강력한 위키백과 생성 도구
* 🤖 스탠퍼드의 Storm AI는 AI 리서치 에이전트를 활용하여 다양한 출처의 정보를 수집하여 **특정 주제에 대한 전체 Wikipedia 문서를 생성**할 수 있는 최첨단 도구입니다.
* 🌐 이 도구는 오픈 소스이며 로컬에서 실행할 수 있지만, **현재 OpenAI API에 의존하는 웹 검색 기능**은 예외입니다.
* 🔧 설치 과정에는 **GitHub 리포지토리 복제, Python 환경 설정, 웹 검색을 위한 OpenAI 및 u.com용 API 키 생성** 등이 포함됩니다.
* 📚 생성된 기사는 **분류학, 진화, 신체적 특성, 행동** 등의 주제를 다루며 **원본 출처를 참조**하는 등 체계적으로 구성되어 있습니다.
* 🖥 스트림릿을 통해 사용자 친화적인 UI를 사용할 수 있어 리서치를 수행하고 **생성된 기사를 볼 수 있는 더 간단한 인터페이스**를 제공합니다.
* 🔍 이 도구는 현재 웹 검색을 위해 **OpenAI와 u.com**을 지원하지만, 향후 **vlm 및 mistol**과 같은 다른 검색 엔진에 대한 지원도 추가할 계획입니다.
* 💬 튜토리얼 제작자는 관심이 있다면 **전체 로컬 설정 가이드**를 만들겠다고 제안하고 시청자가 좋아요를 누르고 더 많은 콘텐츠를 구독하도록 권장합니다.

we have a new project out of Stanford, which is the best AI researcher that I have ever used. Simply give it a topic, and it is going to create an entire Wikipedia about that topic. You can run the whole thing locally, and it works really well. The best part: it will actually reference each of the websites that it got the information from in the Wikipedia itself. That is what you're seeing here. So, I asked it to create me a Wiki based on artificial intelligence, and this is it. It gave me business and industry, healthcare, education, security, basically an entire team of AI agents researched this. Then, for each of the facts that it output, it actually referenced the web page that it got it from. And today, I'm going to show you how to install it, get it all set up, and get it running. And you can run everything locally if you want. The only thing you can't run locally is the web search. So, you can see it's open source. It is running locally on my computer, although we are using the OpenAI API right now.

but you don't have to now. Before I show you how to set this up locally, you can actually test it without setting it up at all. They have a demo page at storm.genie.stanford.edu. And you can see here, they already have pages that are fully researched right here, very very impressive. And one thing that looks like it's coming is this little toggle human AI collaboration mode available soon, which also looks very cool. So now let me show you how to install it. All right, so here's the GitHub page. It already has nearly 62,000 Stars. Hasn't been out that long. Storm stands for Synthesis of Topic Outlines through Retrieval and Multiperspective Question Asking. Now, I did run into one bug and I actually had to change the code to get it working locally, but I'll show you how to fix it. All right, so the first thing you're going to do is open up VSS code. CD to whatever directory that you want this stored at. So I usually like to put

everything on my desktop. Just when I'm first testing it out, so that's what I'm going to do. CD desktop. Next, we're going to come back to the storm GitHub page. We're going to click this little green code button, and we're going to copy this GitHub URL right there. Then we come back to our VS code, and we type git clone. Then we paste in the URL that we just copied, hit enter. That's going to clone that repository to our desktop. Once we've done that, we're going to CD into storm. Next, we're going to use our favorite Python environment management application conda. So conda create DN storm python equals 3.11. Now, I already have an environment called storm, so it's going to ask me if I want to overwrite it, and yes, I do, but you won't see that. Okay, next, I'm going to proceed. Then we grab this command right here, conda activate storm. We're going to paste it in, and then we activate that environment, and we can tell it's active because it

says it right there. Now, from there, we're going to install everything in requirements.txt. So, pip install dasr requirements.txt and hit enter. All right, so that's done. Now, what we're going to do is we're going to go up to file and say open folder, go to the desktop, and we're going to open up Storm. Now, it does reset our terminal, so let's open that back up. And the first thing we're going to do is conda activate Storm, hit enter. So, now we have it active again. And over here, we're going to create a file called secrets.toml, and I'll show you how to do that. So, you're going to right-click new file and call it secrets.toml, and then hit enter. Now, back on the GitHub repo page, we're just going to copy all of this, which is under the getting started header. Copy that code, come over here, and paste it in. Now, the first time going through this, I'm going to show you how to do it with OpenAI, but you can do it through the OpenAI

API, but you can also do it through Azure if you want, but we're not going to be doing it through Azure. So I'm simply going to delete that, delete this line of code as well. So the OpenAI API type is going to be OpenAI. So the first thing we're going to need is an OpenAI API key. So if you don't already have an OpenAI account, go ahead and create one. And then you go to platform.open.com. AI keys. We're going to click this button, create new key. And I'm going to call it Storm YT because I will revoke this key before publishing this video. Grab the key. We're going to switch back to VS code. And we're going to paste it in on line two. And then we can hit save. The next thing we need is a ud.com search API key. And U.com does cost money, but they give you a free trial. So you don't actually have to pay anything if you just want to try this out. So you go to api.u.com dashboard.

It'll ask you to sign in. Go ahead and sign in. We're going to click this create API key button right there. Here under web search, we're going to click generate. And then here is our API key, which I will copy. So, I couldn't actually find a way to revoke API keys from u.com, so I'm not going to show that one. So, we're going to paste in the u.com API key right there and then hit save. And then you're pretty much done. So, there are two ways to get this working. One, you can do everything command line, which is nice and it outlines all of the different options that you have right on this page. But to make it a little bit easier, they created a very simple UI that we can use, and that's my preferred way. It just makes everything much simpler. So, to get that working, you're going to look for this folder called front end. And then within that, there's a streamlet folder. You're going to right-click on secrets. tommel and then

copy it CU we want it to also be in the streamlit folder, and then you're going to right-click on this do streamlit, and click paste, and that will make it available to the front end. Okay, the next thing you're going to do is CD into the demo light directory, then you're going to run streamlit so streamlit run storm. py, and then hit enter, and there we go. It looks like it may have worked on the first try, and here you can see the streamlit server is up and running. We have it on Local Host. It is still trying to spin up, so maybe it didn't work, but we'll see. Okay, so it does look like it worked. Let's start your first research, and then we'll we type dogs, and then we'll do research, and we should be able to see all the logs happening here. So it says brainstorming open it, start identifying different perspectives for researching this topic, and there we go. We can see the HTTP request to the OpenAI endpoint, so it looks like we didn't run into any

bugs this time. Maybe they updated the code since the last time I pulled, which was, I think, yesterday. But good for them. This should work for you. Here we go. It's doing all of the browsing. Look how many different pages it is browsing so fast. Okay, so the article finished. Let's see. Show final article. And here we go. All the information we could ever want. Run locally and it did all the research in real time. This is the most real-time information it could have possibly pulled. So we have taxonomy and evolution, physical characteristics, behavior, all very well organized, and references to each of the articles that it got the facts from. Over here, we can actually select the references and it'll link directly to it and give us a highlight of each of the articles. It is such a good researcher. Okay, but now what if you don't want to hit open A's API? And I guess technically you could probably set up your own web scraper too, although that's a pretty big thing to accomplish. So you can get all of this.

except for the web scraping running locally. If you want, and currently they support vlm with mistol, and they actually have an example of how to do it. But I haven't set it up myself, and I've seen a GitHub issue where they talk about getting o Lama working, but again I couldn't really get it working. I didn't try that hard. If you want to see me do a full local tutorial of this, just let me know in the comments. If you enjoyed this video, please consider giving a like and subscribe, and I'll see you in the next one.

STORM AI: Create Full Wikipedia Pages in Minutes With AI Agents (Tutorial)

Artificial Intelligence (AI), Open Source, Generative Art, AI Art, Futurism, ChatGPT, Large Language Models (LLM), Machine Learning, Technology, Coding, Tutorials, AI News, and more

** Exclusive Pinecone Partner **

Matthew Berman

스탠포드 AI 연구 프로젝트: Storm

- 스탠포드 AI 연구 프로젝트인 Storm은 여러 웹사이트의 정보를 참조하여 다양한 주제에 대한 위키백과 페이지를 생성할 수 있는 AI 연구 프로젝트입니다.

- 로컬에서 실행할 수 있고, 오픈 소스이며, 테스트용 데모 페이지가 있고, 인간과 AI의 협업 모드가 계획되어 있습니다.

- Storm으로 알려진 이 프로젝트는 많은 별을 받은 GitHub 페이지가 있으며, 사용자가 로컬에서 실행하려고 할 때 몇 가지 버그가 발생할 수 있지만 이는 수정할 수 있습니다.

Storm 리포지토리 및 Python 환경 설정하기

이 과정에는 GitHub 페이지에서 Storm 리포지토리를 복제하고 conda를 사용하여 Python 3.

11 환경을 만드는 것이 포함됩니다.

Storm 환경을 활성화한 후 다음 단계는 요구 사항을 설치하고 secrets.

toml 파일을 설정하는 것입니다.

또한 이 과정에는 VS 코드에서 Storm 폴더를 열고 안내에 따라 프로젝트를 초기화하는 작업도 포함됩니다.

OpenAI API를 사용하려면 OpenAI 계정의 platform.

com에서 OpenAI API 키를 받아야 합니다.

com 검색 API도 키가 필요하며, 웹 검색 아래의 u.

com 대시보드에서 얻을 수 있습니다.

두 API 모두 명령줄 또는 간단한 UI를 사용하여 통합할 수 있으므로 쉽게 사용할 수 있습니다.

스트림라이트 설정 및 데모 실행하기

이 과정에는 파일을 Streamlit 폴더에 복사하고 Streamlit 서버를 실행하는 것이 포함됩니다.

이 데모에는 모든 로그와 최종 기사를 실시간으로 볼 수 있는 기능으로 주제를 조사하는 과정이 포함되어 있습니다.

또한 OpenAI의 API를 사용하는 대신 웹 스크레이퍼를 설정하는 옵션도 보여줍니다.

로컬에서 웹 스크래핑 설정의 어려움

로컬에서 웹 스크래핑을 설정하는 것은 특히 VLM과 Mistol을 지원하는 경우 어려울 수 있습니다.

사용 가능한 예제가 있지만 일부 사용자는 GitHub 이슈에서 설명한 대로 작동하는 데 어려움을 겪습니다.

로컬 웹 스크래핑을 성공적으로 구현하려면 추가 지원과 튜토리얼이 필요할 수 있습니다.