00:00:00Andrey Karpathy just gave us the keys
00:00:02to his personal Obsidian RAG system.
00:00:06And I put RAG in air quotes
00:00:07because this Obsidian power knowledge base
00:00:10has no vector database, no embeddings
00:00:12and no complicated retrieval process.
00:00:15Yet it solves the exact same problem
00:00:17that these more complicated RAG structures claim to do,
00:00:21which is allow our large language model
00:00:23to handle large amounts of documents and answer questions
00:00:27and gather accurate information about them.
00:00:30And the best part about this Obsidian powered system
00:00:32is that it is very lightweight, it's essentially free
00:00:36and it is the perfect middle ground
00:00:38for a solo operator or a small team.
00:00:41So today I'm gonna show you
00:00:42how Karpathy's Obsidian knowledge system works,
00:00:45how to set it up yourself
00:00:46and how it differs between traditional RAG systems
00:00:50so you know if this is the right option for you.
00:00:52So the process by which we are going to create
00:00:54this Obsidian powered knowledge system was laid out yesterday
00:00:58in a pretty comprehensive Twitter post by Andre Karpathy.
00:01:02Now the big takeaway from this post
00:01:04is that we are able to create
00:01:05large language model knowledge bases
00:01:07that essentially act in the same way
00:01:09as something like light RAG or RAG anything
00:01:12or any other graph RAG system with Obsidian.
00:01:17And we're able to do so in a rather simple manner
00:01:20by just having a clever structure to our file system
00:01:23and how we actually ingest data.
00:01:25And the end result is that I am able to ingest
00:01:28a pretty significant amount of data and documents
00:01:32into my Obsidian vault and use Claude code
00:01:35to ask questions about it,
00:01:36to figure out connections between different things,
00:01:38AKA the exact same thing you would do
00:01:41with a traditional RAG system,
00:01:43but with none of the overhead and a way simpler setup.
00:01:46And as Andre lays out, the setup looks something like this.
00:01:49First, we have data ingestion.
00:01:51We are bringing in articles,
00:01:52we're bringing in papers,
00:01:53we're bringing in repos from the internet or from wherever,
00:01:57and we're bringing it into a raw directory
00:02:00inside of our Obsidian vault.
00:02:02This is essentially the staging area
00:02:03before it gets turned into a Wiki.
00:02:05We as the human being in this interaction
00:02:07are able to see all of this happening via Obsidian.
00:02:10Obsidian for all intents and purposes is our front end.
00:02:13Here's where I can see where all the documents are laid out.
00:02:15Here's where I can read all the Wikis.
00:02:17So it isn't sort of abstracted away in a black box,
00:02:20like it isn't a RAG system.
00:02:21It's kind of hard, even in a graph RAG setup like Light RAG,
00:02:25to actually go inside of here and really see everything.
00:02:29I mean, I can, but as cool as this looks,
00:02:31this isn't very efficient.
00:02:33And from there, you just do a Q&A
00:02:35via something like Claude code.
00:02:37And like Andre laid out here,
00:02:38he expected that he would have to reach out
00:02:40for something like RAG,
00:02:42but the large language model has been pretty good
00:02:43about auto-maintaining index files
00:02:45and brief summary of all the documents it reads.
00:02:47And this is something we are gonna be able to do too
00:02:49with a pretty simple Claude.md file,
00:02:52which I will be giving you.
00:02:53And you will be able to find that Claude MD,
00:02:55as well as a written guide
00:02:56that comes with a bunch of prompts
00:02:57inside of my free Chase AI community.
00:03:00There will be a link to that
00:03:01in the description of this video.
00:03:03And speaking of Chase AI, and you knew this was coming,
00:03:06quick plug for my Claude code masterclass.
00:03:08Just released this a couple of weeks ago,
00:03:09and it is the number one place to go from zero to AI dev,
00:03:12especially if you do not come from a technical background.
00:03:15You can find a link to this in the pinned comment.
00:03:18So make sure to check this out.
00:03:19If you're serious about learning this tool.
00:03:22Now, before we jump into the specifics
00:03:24of how to set up this Obsidian system for yourself,
00:03:28let's go over the actual file structure
00:03:30because this is important to understand
00:03:32how data is coming into our vault
00:03:34and then getting turned into Wikis.
00:03:36So the Obsidian vault is where everything lives.
00:03:39As you'll see, if you've never used it before,
00:03:41when you download Obsidian,
00:03:42you are going to designate a specific folder as the vault.
00:03:45In my case, it is quite literally called the vault.
00:03:48That's where everything in Obsidian lives.
00:03:50As a sub folder of the vault,
00:03:52we are going to have the raw folder.
00:03:54The raw folder is where all of our research gets dumped.
00:03:58Anything we want to manually include in these Wikis gets put.
00:04:01This is essentially the staging folder.
00:04:02So this is where all the raw data is going to be held.
00:04:05This can be markdown files.
00:04:06This can be PDFs.
00:04:07And I'm going to show you how to use the Obsidian clipper
00:04:10to essentially turn any webpage into a markdown file
00:04:14like it's sent to the raw folder automatically.
00:04:16We will have another sub folder
00:04:18that is the Wiki folder.
00:04:19So what the large language model is going to do,
00:04:21what Cloud Code will do for us, is on demand,
00:04:24or you could have it even be a skill or have it be automated,
00:04:27is we are going to point it at the raw folder and say,
00:04:29"Hey, I want you to create a Wiki about whatever subject
00:04:33"you've been gathering information about."
00:04:35From there, it will then create a Wiki about that.
00:04:37So you can see we have three different Wikis here,
00:04:41one for AI agents, one for RAG systems,
00:04:43and one for content creation.
00:04:45Now, in between the Wiki folder and these sub Wiki folders
00:04:50is the master index markdown.
00:04:53This is essentially just a list
00:04:54of all of the different Wikis that have been created.
00:04:58Because the idea is when you, this is you,
00:05:02when you talk to Cloud Code, all right,
00:05:04that's Cloud Code over there, and say,
00:05:06"Hey, I want to learn more about AI agents.
00:05:08"Can you ask, I want to ask questions about my Wiki."
00:05:12Well, what is it going to do?
00:05:13Well, it's going to go to the vault
00:05:15because you're probably already in there.
00:05:17It's then going to go to the Wiki folder.
00:05:18It's going to go to the master index folder and say,
00:05:21"Hey, what Wikis have we created?
00:05:23"Oh, he wants to know about RAG systems."
00:05:26Okay, it goes down to RAG.
00:05:28And the Wiki folders themselves have index files
00:05:31which break down all of the additional content.
00:05:33So what Obsidian gives us
00:05:35and what this file structure gives us
00:05:36is a very clear path to find information,
00:05:39even if we have a ton of it floating around.
00:05:41And this helps Cloud Code
00:05:42because it's not going to have a ton of issues
00:05:45finding the data.
00:05:46We're not going to run a million tool calls
00:05:48to see what's in our file structure.
00:05:50But it also helps you because it's very clear where to go.
00:05:52For example, over here on the left is my Obsidian folder.
00:05:56I'm in the Obsidian UI,
00:05:57and we'll go through the download here in a second.
00:05:59But if I want to see a Wiki, what do I do?
00:06:01I just go to Wiki.
00:06:03I have a master index
00:06:04which lays down everything in there.
00:06:06Right now, it's just three things.
00:06:07But if there were 3,000, it still wouldn't be too difficult.
00:06:10And then from there, you know, I can click on it.
00:06:12It takes me to the index of that specific Wiki.
00:06:16And then I can look at different stuff inside of there.
00:06:18It's that simple.
00:06:19And it's that simple for AI too,
00:06:21which is why we're able to use
00:06:22essentially just a markdown file structure
00:06:24to somewhat mimic a rag system.
00:06:27So while that theory is cool,
00:06:28now let's go into how to actually set this up for yourself.
00:06:31First and foremost, you're going to need to download Obsidian.
00:06:33You're just going to head to obsidian.md, hit Download Now,
00:06:37go through the wizard.
00:06:38It's completely free.
00:06:40And you're going to designate some folder as the vault.
00:06:43Just create one, call it the vault.
00:06:45Makes it easy for me, and it'll probably work for you.
00:06:47After we create the vault,
00:06:49we now need to set up this file structure inside of it.
00:06:52The easiest way to do that is with Clawed Code.
00:06:54Simply open up Clawed Code in the vault.
00:06:57So that's the directory I'm in.
00:06:59And you're going to give it a prompt
00:07:01telling it to create this file structure.
00:07:03Now, luckily for you, I already created the prompt.
00:07:05So you can just copy this thing and paste it in a Clawed Code.
00:07:08Now, if you're like me and you've already been using Obsidian
00:07:10for a bit, you probably have a bunch of folders
00:07:13already in there.
00:07:14So maybe you don't want to call it raw.
00:07:17Maybe you want to call it something else.
00:07:18The whole point of it is you just need to designate
00:07:20some folder as, like I said, sort of the holding area
00:07:23or the staging area for where all this information
00:07:25is going to get dumped until it gets turned into a Wiki.
00:07:27So just as needed.
00:07:28Now, the next thing we want to do is create a Clawed.md file.
00:07:31Personal assistant type projects, things like this
00:07:33that are very Markdown heavy, Clawed.mds are perfect for.
00:07:37And this Clawed.md file is breaking down
00:07:40the knowledge base rules,
00:07:41as well as how to essentially traverse it.
00:07:43So again, that we aren't wasting tokens
00:07:44when we ask questions.
00:07:46Again, I have this entire Clawed.md template prompt
00:07:50you can use.
00:07:50This Clawed.md file is also telling Clawed
00:07:53how to structure these Markdown files.
00:07:55So it's very easy to traverse files
00:07:58with this Wiki links format.
00:08:00Now let's talk about how we can bring things
00:08:02into this raw folder.
00:08:03How we can get data into our system in the first place.
00:08:06Well, super easy way to do this
00:08:08is with the Obsidian Web Clipper.
00:08:10So I will put a link to this in the school,
00:08:13or you can go to obsidian.md/clipper.
00:08:16And this is just a Chrome extension,
00:08:18which makes it super easy to turn a webpage into data,
00:08:22into a Markdown file.
00:08:23Now, the one issue with this Web Clipper
00:08:25is it's going to struggle with images.
00:08:26It's just not even going to bring them in.
00:08:27I'll have them as a link.
00:08:29But I want to be able to see the images from these documents
00:08:31I ingest inside of Obsidian.
00:08:33So what do we do?
00:08:34Well, we are going to use an Obsidian community skill
00:08:37or Obsidian community plugin to help with this.
00:08:39So one of the cool things about Obsidian
00:08:41is the community plugins.
00:08:42There's thousands of them.
00:08:43So if you're inside of Obsidian,
00:08:46I'm inside the desktop app right now.
00:08:47If I come down here and I hit this little gear,
00:08:50I'm going to go to community plugins.
00:08:52I'm going to go to browse.
00:08:54And then you're going to search for local images plus.
00:08:56You're going to download it, install it, and turn it on.
00:09:00Make sure it's enabled.
00:09:01You can confirm it's enabled
00:09:03by heading to your community plugins tab
00:09:05and seeing this little tab turned on.
00:09:08Now, if we use the Obsidian web clipper,
00:09:11and I can see that over here as an extension,
00:09:13you can see what happens.
00:09:15It immediately pulls everything.
00:09:17And if I hit add to Obsidian,
00:09:19I can see this entire article, including the images.
00:09:21Now there is one thing we need to set up
00:09:24inside of the web clipper,
00:09:25and that's making sure it actually pulls it
00:09:26into the raw folder automatically.
00:09:29I don't want to have to manually do that.
00:09:30You're just going to head to the options on your web clipper.
00:09:34I just right clicked it.
00:09:35And then over here on the left, where it says default,
00:09:38I created my own new template,
00:09:39but you can stick on the default if you want,
00:09:42where it says location and note location right here.
00:09:47You're going to want to change that from clippings to raw.
00:09:52And that will make sure when you use the web clipper,
00:09:54it automatically goes into the raw folder.
00:09:56So now with the Obsidian web clipper extension
00:09:59and the images community plugin,
00:10:01we can now turn any webpage on the internet
00:10:04into a markdown file that will be used for our Wiki.
00:10:08But that is just one data funnel.
00:10:10That's a manual one.
00:10:11We can have Claude code do a bunch of heavy lifting too.
00:10:14So let's say I was trying to create a Wiki
00:10:16about Claude code skills.
00:10:17So I told Claude code,
00:10:18let's create a Wiki about Claude code skills.
00:10:20I already included some info in the raw folder,
00:10:23what we pulled in via the web clipper.
00:10:25Go conduct your own research and bring in the relevant raw
00:10:27MD files to generate that Wiki.
00:10:29So what is it going to do?
00:10:30It's going to go on the internet, use its standard web search,
00:10:32and it's going to create its own Wiki about Claude code
00:10:36skills.
00:10:37So what you see is that this raw folder,
00:10:40this whole raw pipeline, that's more for you.
00:10:42That's for when you mainly want to put in some information.
00:10:44Now you can have Claude code do that as well,
00:10:46but Claude code is also smart enough to essentially take the
00:10:49research,
00:10:50figure out what's relevant itself and just create the Wiki
00:10:53directly. This raw folder is really for you,
00:10:55the human being to have some level of organization.
00:10:58And here's what Claude code came back with.
00:10:59So it created the Claude code skills Wiki.
00:11:02We see here in the master index that it's referenced here.
00:11:05If I click on it,
00:11:07this then brings us to the index of Claude code skills.
00:11:10And right now it has four articles.
00:11:12So here's the skills overview article.
00:11:15You can see it links to websites and it also links to
00:11:18different articles within our obsidian vault.
00:11:21So if I click on skill ecosystem, here's more stuff.
00:11:25I click on the top skills, right? So on and so forth.
00:11:27There's a very clear pathway from one article to another and
00:11:30how these things relate,
00:11:32which means when you ask Claude code questions about these
00:11:34articles in these subjects,
00:11:35it's easy and cheap for it to answer questions about them,
00:11:39which then brings us to the obvious question.
00:11:41Do we need rag at all? You know,
00:11:43we look at something like this light rag setup.
00:11:45You watch my last few videos with light rag and rag,
00:11:48anything, and seeing how simple the set up with obsidian,
00:11:51you're probably like, well,
00:11:52why would I ever even bother with these more complicated
00:11:55setups at all?
00:11:56And the truth is if you're a solo dev,
00:11:59a solo operator or a small team that isn't dealing with
00:12:02thousands of documents,
00:12:04the answer probably is obsidian makes more sense for you.
00:12:08It's lightweight and you really don't need rag.
00:12:11These large language models,
00:12:12these harnesses like Claude code are good enough for your use
00:12:16case.
00:12:17And we can sit here and get in the weeds about the
00:12:18differences between the obsidian rag and true rag.
00:12:21But the truth is the big thing is scale, right?
00:12:24Are we trying to scale to millions of documents or are we not
00:12:27because at a certain scale,
00:12:29it's going to be cheaper and faster to use a proper rag
00:12:32system.
00:12:33No matter how good Claude code is at navigating this MD file
00:12:38document network you've created.
00:12:40But this isn't a question you necessarily need to have the
00:12:42exact answer to right away.
00:12:44Why wouldn't you just start with something like obsidian?
00:12:47And if it's clear,
00:12:48your scale goes well beyond the bounds of what this thing can
00:12:51handle, then just move into rag.
00:12:53I think people get really caught up in like answering this
00:12:55question when it's like, just try it out, just experiment.
00:12:58It's not costing you anything to use some sort of rag system,
00:13:01rag system like obsidian.
00:13:03And if it doesn't work, it doesn't work fine.
00:13:05Then go to use light rag and said,
00:13:06people want to sit here as they inevitably will in the
00:13:09comments and like argue this back and forth, just try it.
00:13:11I think the answer will be pretty clear at a certain point
00:13:14when you need to move to a true rag system.
00:13:16But the nice thing with this is, is again,
00:13:19most people don't need a real rag system.
00:13:21They just don't, right?
00:13:22Even if they're in a small business team situation.
00:13:24So having a proper, you know,
00:13:27orchestrated system like the subsidiary and knowledge base,
00:13:30I think is a huge boon to the majority of people.
00:13:33So I hope this breakdown was useful to you.
00:13:35Definitely check out Andre's post about this.
00:13:37He goes into a fair amount of detail.
00:13:39Make sure to check out the free chase AI school.
00:13:41There's a link to that in the description that has all the
00:13:43prompts and a written breakdown of how to actually do this.
00:13:47If you got confused at any part and as always take a look at
00:13:50chase AI, plus if you want to get your hands on that
00:13:52masterclass besides that,
00:13:54let me know what you thought and I'll see you around.