Karpathy LLM Wiki: Build It With Obsidian and Claude Code
Andrej Karpathy posted a short tweet last week that made about half the personal RAG stacks on the internet look like massive overkill. No vector database. No embeddings. No retrieval chain. Just a folder of markdown files, Obsidian, and one schema file that Claude Code reads every session. He is running a 400,000-word personal research wiki on it.
The LLM wiki pattern swaps retrieval for compilation. Claude Code reads each new source once and files it into a markdown wiki it maintains, following one CLAUDE.md schema file. Below about 500 articles it beats vector RAG on setup time, cost and how easy the knowledge is to browse.
This post is the complete walkthrough of the pattern and the exact schema file every other tutorial glosses over, so you can build the vault yourself today.
The Core Insight
Most people's experience with LLMs and documents looks like standard RAG. You upload a collection of files, the model retrieves relevant chunks at query time, and generates an answer. This works, but the model is rediscovering knowledge from scratch on every question. There is no accumulation. Ask a subtle question twice, the model has to find and piece together the relevant fragments twice.
The wiki pattern inverts this. Instead of retrieving from raw sources at query time, the LLM incrementally builds and maintains a structured wiki that sits between you and the raw sources. When you add a new source, the model reads it, extracts the key information, and integrates it into the existing wiki. Updating entity pages, revising topic summaries, noting contradictions, strengthening the synthesis. The knowledge is compiled once and then kept current, not re-derived on every query.
The wiki is a persistent, compounding artifact. The cross-references are already there. The contradictions have already been flagged. The synthesis already reflects everything you have read. It gets richer with every source you add and every question you ask.
The Three Layers
Every version of this pattern has the same three layers. Get these right, nothing else matters.
Layer 1: raw/
This is your inbox. Articles, papers, transcripts, pasted notes, screenshots. The LLM reads from this folder and never writes to it. Everything in here is immutable. It is your source of truth.
Layer 2: wiki/
This is where the LLM lives. It writes summaries, entity pages, concept pages, topic indexes, and the master index. You read this layer in Obsidian. You do not write in it. Manual edits cause the system to drift session over session.
Layer 3: CLAUDE.md
One file at the vault root. The schema. This is what turns a generic Claude Code session into a disciplined librarian. Every session reads this file first. Every operation follows rules from this file. This is the piece every other tutorial skips.
Optional fourth layer: output/ for query results, reports, slide decks, and generated artifacts. Good outputs get promoted into the wiki as new articles so your explorations compound.
The Four Operations
Four verbs run the whole system.
Ingest. Drop a source in raw/ and type compile in Claude Code. The model reads the file, decides which topic it belongs to, writes or updates a wiki article, adds backlinks, updates the topic index, updates the master index, and appends a log entry. One source typically touches 10 to 15 wiki files in a single pass. That is the bookkeeping humans abandon and the reason every second brain system eventually dies of neglect.
Query. Ask a question. The model reads the master index first, then the topic index, then 1 to 3 specific articles. Three to four file reads, no vectors, no embeddings. The index files are the retrieval layer, and the model maintains them for you automatically.
Lint. Periodic health check. The model reads every file in the wiki and produces a report covering contradictions, stale claims, orphan pages, missing cross-links, unsourced claims, and suggested new articles. No changes happen during the lint pass. You approve specific fixes one at a time.
Log. Every operation appends one line to wiki/log.md. Append-only, timestamped, parseable with grep. Gives the wiki a memory of what happened when.
The CLAUDE.md File (The Part Nobody Shows)
Every walkthrough online calls this file "the brain" and never opens it on camera. Here is the exact structure that makes the pattern survive across sessions.
The file is about 60 lines. It opens by telling the model who it is:
> You are the librarian of this vault. The wiki/ folder is your domain. You write and maintain every file in wiki/. The human rarely edits wiki files directly.
That sentence is load-bearing. Without it, Claude Code treats wiki files like any other files and starts deferring to whatever it sees there. With it, the model takes responsibility and rewrites pages confidently when new sources conflict with old ones.
The next section defines the four operations with numbered procedures. When you type compile, the model runs seven specific steps in order. When you ask a question, the model runs five specific steps. When you type lint, the model produces a seven-section report. These procedures are what makes the behavior consistent across sessions.
The conventions section enforces the rules that keep the wiki honest:
- Every wiki article cites the raw source file it was compiled from.
- Every article includes a Key Takeaways section.
- File names use lowercase with hyphens.
- Wikilinks are required for every cross-reference.
- Bullets over paragraphs.
- Never invent claims. Flag gaps in an Open Questions section.
The citation rule is the hallucination fix. Every article has to name the raw file it came from. If the model writes something the source does not say, the next lint pass catches it. That is how you keep a wiki honest at 200 articles.
The setup steps for the full vault are at the bottom of this post.
Live Query Walkthrough
Three query patterns work well against a wiki built this way.
Direct lookup. Ask a specific factual question about a single article. The model reads the master index, finds the topic, reads the article, answers. Three file reads.
Cross-topic synthesis. Ask a question that spans multiple topic folders. The model reads the master index, then multiple topic indexes, then 2 to 4 articles across topics, synthesizes. Six file reads for a complex question.
File-back synthesis. Ask a question and tell the model to file the answer back into the wiki. The model produces the answer as a new wiki article in the appropriate topic folder. Your exploration compounds into the knowledge base. Every future query benefits from this answer being present.
All three patterns run in seconds. None of them touch a vector database.
Vector RAG vs the Wiki Pattern
The question everyone asks is whether this replaces vector RAG. The honest answer is below 500 articles, the wiki wins on four out of five factors.
| Factor | Karpathy Wiki | Vector RAG |
|---|---|---|
| Infrastructure | Folder of markdown files | Vector DB, embedding model, hosting |
| Setup time | 15 minutes | Hours to days |
| Scale ceiling | ~500 articles | Millions of chunks |
| Human browsable | Read and navigate freely | Black box |
| Outputs compound | Queries file back as new articles | Chat is ephemeral |
Vector RAG wins on scale. Wiki wins on everything else. Below 500 sources, the wiki is strictly better for solo operators. Above that, hybrid makes sense: wiki for structured synthesis, vector for semantic fallback across long-tail retrieval.
Scale Ceiling and Hallucination
Two objections come up every time this pattern gets posted.
Scale. The wiki starts breaking down around 500 articles because the master index stops being a reliable navigation layer. The fix is either to split into multiple topic-specific vaults with a top-level router, or to bolt on a small BM25 or hybrid search tool over the markdown files. Karpathy mentions using a local search tool at larger scale. You probably will not hit this ceiling for a year.
Hallucination drift. If the model writes a wiki page that drifts from the source, the error propagates into every future query. The fix is in the schema file above. Every article cites its raw source. Every lint pass checks citations against sources. Flag anything unsourced, review it, correct it. This is not automatic safety. It is a maintenance loop the model runs for you.
Build the Vault
A working vault needs five things: the raw/ and wiki/ folders, the CLAUDE.md file with all four operations, three example pages (entity, concept, source summary) so the model has shape references, a starter master index and a starter log.
Create the folders, write the CLAUDE.md from the structure above, and open the folder as a vault in Obsidian. Run Claude Code in the root. Drop a sample article in raw/. Type compile. You have a working self-maintaining wiki in minutes.
New to running Claude Code in a folder? Start with Claude Code vs regular Claude. The same folder-as-memory idea runs my Claude Code daily brief, and Claude Code skills examples shows how to save compile and lint as skills.
What Comes Next
The wiki pattern is the foundation. Once it works, the useful extensions start showing up:
- Multi-vault federation. One vault per domain, a top-level CLAUDE.md that routes queries across them.
- Auto-ingest. A cron job that watches a Gmail label, a Slack channel, or an RSS feed and drops new items into
raw/automatically. The model processes on schedule. - Agent handoff. Separate agents with different roles all reading from the same wiki. Research agent ingests, executive assistant queries, content agent publishes.
- Voice note ingestion. Record a voice memo, transcribe with Whisper, drop into
raw/, the model files it as a dated journal entry cross-linked to relevant concepts.
All of these reuse the same vault template. The pattern compounds.
Key Takeaways
- The wiki pattern compiles raw sources once into a persistent structured artifact instead of retrieving from them on every query.
- Three layers: raw (immutable source of truth), wiki (LLM-owned synthesis), schema (CLAUDE.md configuration).
- Four operations: ingest, query, lint, log.
- The CLAUDE.md file is the load-bearing piece every tutorial glosses over.
- Below 500 sources the wiki beats vector RAG on almost every axis.
- Build the folders, write the schema, open it in Obsidian, run Claude Code and start compiling.
Free community: https://www.skool.com/stride-ai-academy-7057
Transcript
Everything you're looking at right now was built 100% with Claude Code and Obsidian. Every wiki page, every backlink, every summary, I simply dropped in raw articles, raw transcripts into a folder and walked away. When I came back, I had this. Now, this entire pattern didn't come from me. This came from a Andre Karpathy tweet, which went mega viral, and this whole entire setup takes about 15 minutes, and you're going to be getting access to this finished vault template at the this video 100% for free, no paywall or anything like that. Now, you've probably saved articles, podcasts, notes, whatever the case is, you know, in Notion, PDFs, podcast transcripts, and you meant to review these, but you never got around to them, and part of the reason of that is is because none of it is actually searchable. None of it's connected, it's just scattered all over the place, and none of it actually feeds into your work. And every second brain system you may have tried out there, whether it's Notion, different systems and setups, they typically fail for the same reason. You stop maintaining it, and the reading and thinking was never really the problem, the bookkeeping was. Now, Andre Karpathy just posted the fix on Twitter, it went viral, he posted a gist as well, which really dives deep, and all this will be linked, of course, in the description down below. I'm also going to be giving you a free resource, it's a 12-page document that literally goes over everything, goes over in-depth this entire system from start to end, which we're, of course, going to dive deep into in today's video. I also will be giving you this exact template, so you can literally just copy and paste and start using this by the end of this video for your own knowledge system. Now, before we dive in, just to give you some quick context in case you don't already know, who is Andre Karpathy? Well, he was a part of the founding team at OpenAI, he was the head of AI at Tesla, built the entire autopilot vision stack, he created one of the most popular deep learning courses on the internet. You may have seen his recent GitHub project go viral, auto research. And when this guy posts a workflow for managing knowledge, or really anything about AI that he posts, it's worth paying attention to, and really everyone just listens. So, currently he is running a 400,000 word personal research wiki with no vector database, no embeddings, no retrieval chain, just markdown files and Obsidian, and one schema file that Claude Code reads every single session. Now, in the next 10 minutes or so, I'm going to show you the entire pattern so you have a deep understanding of it, every line of the schema file that nobody else wants to explain. I'm going to show you a live compile, a live query, a live lint, and by the end of this video, you will have a working system and a template that you can run on your own stuff. Now, every version of this pattern has the same three layers. Get these right and nothing else really matters. So, layer one is raw. This essentially is your personal inbox as a human. So, what goes in here are articles, papers, transcripts, screenshots, anything that you dump in here that you want the LLM to read, and that's where it's going to read is from this raw folder. Now, keep in mind the LLM is never going to write into this folder. This is literally just an inbox for you to dump things in, and this is immutable. It is your source of truth. Now, layer two right here is the actual wiki, so this is where the LLM files live. It writes summaries, it writes entity pages, it writes concept pages, it builds an index, it maintains cross-links between every article. You read this layer in Obsidian. Now, you as the human do not write in this layer. Not because you can't, but because every edit that you make in here is one that the model cannot predict in the next session, and the whole system starts to drift. And then at the top here for layer three is the schema. This is one claude.md at the vault root. And this is what turns a generic Claude Code session into a disciplined librarian. Every session it reads this file first, every operation follows rules from this file, and this is the core piece of the system. Now, if you want a deep dive into claude.md files, make sure to check out this video right here I did a few days ago on claude.md files and how to properly structure them, but of course, like I mentioned, I'm going to be giving you this free template, which includes a structured Claude MD that you can use out of box. So, raw, wiki, schema. That is the whole architecture. Now, let's dive into the actual layers and how this all works. So, there's really four operations, and we're going to start off with the first one, which is ingest. So, this is where you drop one source in the raw folder, and you tell Claude Code to then compile it. And in one pass, it can touch 10 to 15 wiki pages. Here's what it actually does though under the hood. So, it reads the raw file in full, then it runs checks in the master index to see if the topic already exists. If yes, it updates the existing article and adds backlinks from any related pages. If no, it creates a new topic folder with its own index file. Either way, it updates the master index and appends a line to the log, and the wiki has just compounded by one source. All right, so let me show you how this actually works in action. So, first things first, you're going to want to download the Obsidian web clipper. So, link to this will be in the resource below, and this whole document, as well as all the different resources, templates, etc. from this video and others is available in our free Stride AI Academy. Now, once you go ahead and actually download this, you'll see the Obsidian icon right here for the extension. We can go ahead and click on it. You're going to want to click on settings to open up the Obsidian settings. You'll see right here, I actually selected the name of my specific vault. By default, it will save it to the open vault. You can also create new templates for how it's going to save it, or use this default one right here, and all you're going to want to do is just change this note location. It's going to be clippings by default, but you can change this to raw, or whatever you want to have your actual intake folder to be. Once we have that set up, we're going to find a blog or whatever piece of information that you want to actually ingest into the system. For this, we're going to be using one of Claude's blog posts right here. I'm just going to go ahead and click on this, and then I'm going to click add to Obsidian. Next, it's going to ask me to open Obsidian. And boom, here you can see we have this entire article with different properties, such as the title, the source, the author, when it was published, created, a description, any specific tags. And you can see here, it literally pulled in everything, including the images. Now, by default, it won't actually pull in the images, but we're using a plugin right here called local images plus, which by default is actually installed in the template that I'm providing you for free. Actually, every single plugin that I cover is actually installed in this template by default. But if you already have an Obsidian setup and you're setting this up in there, or whatever the case is, maybe you just need to install this plugin yourself. You can just go over to community plugins, make sure you turn them on, and then you're going to want to search over here, browse, and you're going to want to search for the specific plugins that we cover. So, I have DataView installed here. I also have local images plus, as well as terminal. So, the terminal one right here, you can actually use Claude Code if you want in Obsidian if you don't want to leave uh Obsidian whatsoever. I personally usually prefer using it in something like Cursor's terminal or VS Code, and I have Obsidian and my actual IDE open at the same time. So, here you can see we have that same Claude blog post that we just saved into our Obsidian, and I'm going to say to Claude, compile, and then I'm linking to that specific uh blog post right here, that markdown file, compile this one into the wiki. That's all I'm going to say, and Claude's just going to do its magic. It's going to take maybe uh 30 seconds to a minute, depending on how much you're compiling, and it's going to actually go about the process. So, watch what happens. It's going to read the article, it's going to identify the core topics, going to decide what needs its own concept page. If the topic doesn't exist yet, it's going to then write the summary, it's going to write a key takeaway section, it's going to add inbound and outbound wiki links, and then it's going to update the topic index, and it's also going to update the master index, and then it's going to log the entry. All from one simple command, which, of course, is powered by our claude.md file. So, before Claude Code, Obsidian was kind of like a big scary tool for a lot of people because you have to do all these different things, backlinks. It was great for bookkeeping your notes and knowledge, but it's something that humans aren't really going to do, and they're just going to abandon, you know, 15 edits across eight files from one source, you would never personally do this manually, but Claude Code does this for us in 30 seconds, and it's easy as that. So, if you start building this up, you do it 10 times, you start to have a real knowledge base, and if you do it maybe like 100 times or a couple hundred times, you know, the graph view is going to start looking like actual research. And boom, it's done. You can see we have a new topic, which is agent design patterns with five articles. So, we have three key patterns, composable general tools, progressive context management, prompt caching strategies, and declarative tool designs. You can see it also updated the master index, the log.md, plus added backlinks with a total of 11 files touched. And like I said, the template that I'm providing you for free with this entire knowledge system comes with a complete claude.md file, and this is really what makes this system work. But you can, of course, customize this if you want to change the system for your specific flow. You can see here, it starts off saying, "You are the librarian of this vault. The wiki {slash} folder is your domain. You write, you maintain every file in this wiki. The human rarely edits wiki files directly." Now, this basically is defining the ownership. Without it, Claude Code treats wiki files like any other file and starts deferring to whatever it sees there. But with it, it's going to start taking responsibility and write these pages confidently with new source conflicts. So, you can see here for ingest, read the raw file in full, identify the core topic or topics, check wiki master index for matching topic folder. If the topic exists, update or extend the relevant articles and add backlinks from touched pages. If the topic does not exist, create a new folder under wiki with a lowercase hyphenated name and create a underscore index.md inside of it. Every wiki article must include a top level title, source which is path to raw md file line, a two to four sentence intro paragraph, a key takeaway section with bullet points, a related section with wiki links to three to eight related pages, update the topic underscore index.md, update the wiki master index if a new topic was created, append one line to wiki.log.md if the source spans multiple topics create articles in both and cross link. We have about 10 steps here and the model's going to follow them in order every single time because this file in the cloud.md tells it what to do and it's loaded in every single conversation so you won't have to prompt it again. Next we have the query section so this is triggered when the human asks a question so read wiki master index first then read the matching topic index.md, read one to three specific articles in full, synthesize the answer with citations, and then offer to file substantial answers as new wiki articles. So three to four file reads to answer any question. No vectors, no embeddings, no cosine similarity, BM25, the index file is the retrieval here and that's because the model maintains it for you. All right, so next is the lint which is the health check. So live is the append only record and we're going to run both the query and the lint live in just a second and you can see here this is going to be triggered simply by just saying the word lint or audit the wiki. It's basically going to read every file in the wiki and produce a report covering contradictions, stale claims, orphan pages, missing concepts, missing cross links, unsourced claims, and suggested new articles and then output the report to output-lint-report and then the date here and then wait for approval before fixing. And then for log every operation appends one line to the wiki.log.md in this time, operation, short description, and then the files touched. And this is append only it never rewrites existing lines. And then we have conventions so things like citations are required every wiki article includes source line, key takeaway section is required on every article, file names are lower case hyphenated, use wiki links for every cross reference, bullets over paragraphs, never invent claims, flag gaps and open questions, and then flag contradictions when found. And then when the human asks something outside of the rules ask a clarifying question, do not silently invent a new operation. So the citation rule right here is the hallucination fix. Every article has to name the raw file it came from. If the model writes something that the source doesn't say that in the next lint pass catches it. And that's how we're going to keep this wiki honest at scale. And that's literally the entire system. It's essentially a 60 line or so cloud.md markdown file. And this essentially is the brain of the entire system and that's why I went through it for you guys because it's very important for you guys to know it. Now just so you guys know in our template we do have some additional things that Karpathy doesn't even cover and that I just felt did add value to this system so I'll quickly go over them. We also have an AI-research folder. So this is the AI's research folder. So you're going to see in a second yes we can ingest manually through the means I just showed you with the Obsidian Clipper or just you know scraping our own podcast transcripts or whatever the case is from different sources. You can also get Claude code of course to conduct autonomous research on the web and save the full clean source content into this folder as markdown files. And you'll see here we're telling Claude that it can write to this folder. And files here are immutable once saved, do not overwrite, create new files. And this separates human curated sources which are in the raw folder from AI discovered sources which are in the AI research folder. And you'll see research is triggered when the human asks you to research a topic or when a query reveals gaps the wiki cannot answer from existing sources. What it's going to do is search the web for relevant high quality sources on this topic and then for each source found save the full clean content not a summary as markdown files in this folder. You can see the format here and we're basically just giving it some additional stuff for the format that it's saving it as. So in the doc here if you want to see more in depth stuff such as the folder structure, the page templates, entity page templates that we have within the system as well as the concept page template, the source summary template, you can see all that there. But we're just going to move on to the query pattern. So there's really three types of queries that work well against the wiki built this way. The first is the direct lookup. We have direct lookup, cross topic synthesis, and file back synthesis. I'm going to run all three against this vault that already has about 10 plus articles as well as transcripts from some of my YouTube videos inside of it. All right, so I'm actually just going to use the Obsidian terminal right here but you can use either or the IDE with cursor, VS code, whatever the case is. I'm just going to say what are the key points from my cloud.md video. And you can see here I have my YouTube transcripts which includes the cloud.md video. So this is a direct query. You can see here let me find the relevant raw files and query the wiki for additional coverage. There's a wiki article specifically about cloud.md best practices. And boom here you can see this is exactly what I covered in that video. If you didn't haven't watched that video I definitely suggest you to go watch it. We covered two different research papers right here. Why Claude ignores your instructions, the fix so six different sections right here, and then some practical tips. All right, so that's the first query direct lookup. Next is the cross topic synthesis. So I'm going to ask what techniques are mentioned across multiple videos. You can see here nine videos across 11 topics. Let me read the topic indexes to trace which videos feed into which articles then identify overlapping techniques. And boom here we got our answers. So we can see technique number one context/token management so that's from videos one, two, three, four, five. And we can see what it pulled there and then we can see technique two prompt caching from videos one, three, and four. And then we can see number three composable general tools over task specific tools video three, four, and nine. And then dead weight pruning, reevaluate assumptions videos two and three, and then cost monitoring and mitigation videos one, five, and nine, security boundaries, and then persistent always on agents. All right, so query three file back synthesis. So this is where the answer becomes a new wiki. Basically the cool thing about this is the exploration compounds into new knowledge in your knowledge base. So here we're saying based on everything in the wiki what are the main approaches to long-term memory for AI agents. Save your answer as a new wiki article and cross link in the sources you referenced. And boom you can see here it is now done so here's what it created a new article right here. It's under the wiki agent design patterns and we have long-term memory approaches so you can see this whole entire document right here. You can see it has links to related documents right here so you could see multbook VPS setup, AI workflow builder, cloud.md best practices. So we can see it synthesizes four approaches from across this wiki so memory folders, compaction, structured config files, external rag. The overarching finding simpler model native memory is replacing external infrastructure as models get more capable which we're kind of seeing right here in front of our eyes with this system. Now this query is great because everything you ask adds value to the system. You never lose a good answer to chat history like you may have in the past and this is the part that turns a knowledge base into a research partner. All right, so next we're going to do the AI research query. So this is basically a query that the AI shouldn't have access to within the current knowledge base. So I'm going to ask it research what QMD is and how it works as a research layer on top of markdown wikis. Save what you find to AI research. You can see here it is doing its research. It's running the fetch multiple different times for different things right here. If you don't know about QMD I'll dive deeper into this in future videos. We do mention it in the resource down below. This is a tool for BM25 rag which is created by Toby who is the founder of Shopify. Okay, and boom you can see that it saved the different research in our AI research folder right here and you can see over here in our knowledge search wiki we have QMD right here and we can see all the different stuff about it BM25, vector semantic, hybrid plus re-ranking. Really cool stuff and just to go over one more time each wiki has its own index so we can see all the different stuff within this wiki right here. So QMD overview, indexing and chunking, MCP integration. And we have that for every single wiki with its own index and then we have the master index which links to each specific wiki. So this is sort of like progressive disclosure. This is exactly progressive disclosure like we talk about in some of our other videos. I talked about it in the cloud markdown video as well. This is how skills work. We're giving it a short description right here for each specific wiki and then it's able to distinguish which wiki it should actually go to based on this description. And then once it's in this wiki we can see as well another index for the specific wiki where it has a description for each specific article. All right, so at this point if you stopped right now you'd have a full working system but in the next few minutes I'm going to show you how to actually keep this system healthy at 100, 200, or 300 articles. So now all I'm going to do is just simply run lint. And once again the lint operation it basically reads every file in this wiki, cross references them, and returns a report. So contradictions, stale claims, orphan pages, missing cross links, unsourced claims, all these things that you know humans as us may miss this is going to be able to detect. And now we can see it's going to work. All right, and boom the lint is complete so we can see our report is saved right here in the output folder. So we can see here we had no contradictions, no stale claims, no orphan pages, but we did have six missing cross links, three unsourced claims, and then five suggested articles. And in this report, we can see all the different details for each specific one. And then we can see here it's waiting for our approval before it actually applies those fixes. So I'm going to go ahead and say I approve. All right, and boom, we can see that all the fixes are applied. So run this weekly or after any big ingest batch. This will take 5 to 10 minutes, give or take, give you a healthy wiki and a log entry so you can know when you last ran it. Now the question everyone's been asking is whether this actually replaces Vector Rag. And the honest answer is below 500 articles or so, this wiki wins on four or five factors. So on infrastructure, Vector Rag needs a database and an embedding model, you need hosting. It's a lot more complex where the wiki is literally just a folder of markdown files. So the wiki wins in terms of simplicity and setup time. Vector Rag may take a few hours to maybe even a couple days depending on how complex the setup is. So the wiki wins with literally just a 15-minute setup. You can literally set it up and have it fully running by the time you're done watching this video. Now for scale, there is a ceiling. So Vector Rag handles millions of chunks. The wiki starts breaking down past maybe 500 or a few thousand articles. So in that case, Vector is going to win. And this here is one of the axes that actually matters. Next for human browsability, Vector Rag is literally just a black box. So the wiki is a set of pages that you can read and navigate and you can view in a nice GUI interface like Obsidian. So in that case, the wiki wins. And the cool thing is with the wiki, outputs compound. With Vector Rag, the chat is ephemeral. So wiki queries file back as new articles. So in this case, the wiki would win. So four out of five below 500 sources, the wiki wins. Above that, a hybrid approach makes sense. So wiki for structured synthesis and Vector for semantic fallback across long-tail retrieval. Now this could change as models progress and whatnot, but that's also why I reference QMD which you can definitely take a look at and I'll do more videos talking about QMD and showing you guys how to actually use it, but it is linked in the resources well. Now you could also use other vector or embedding strategies besides QMD, but that's just one I'm referencing here. There's a few objections that have been coming up every time people have been posting this on X and whatnot and let me handle some of them directly here for you guys. So the first objection is scale. So we talked about this already, but just as a baseline, the wiki starts breaking at around 500 articles because the index stops being a reliable navigation layer. Similar to when you bloat Claude Code with too many different skills, you will start to see some degradation in quality. Now above that, you either want to split it into multiple topic specific vaults or you would want to bolt on a small search tool over the markdown files. Like I said, I'm linking to QMD which is the same search tool that Karpathy actually uses, so that will be in the doc. And you probably won't hit a ceiling for quite a while depending on what your usage is with your actual vault, but then you may want to actually scale it and look at some of these other methods. So either a different vault or a hybrid strategy using something like QMD. The next objection is hallucinations. This is huge when dealing with really anything with AI models. So if the model writes a wiki page that drifts from the actual source, the error propagates into every future query. So it's essentially like a virus. So the fix for this is that Claude.md file and that's why we took a deep look at it so you have an understanding how this actually works. Now every article sites the raw file that it came from. So every lint pass that you do in future, checks those citations against the actual sources. Flags anything that's unsourced, it reviews it, corrects it. Now this is not automatic safety, but it is a maintenance loop for the model that runs for you. And like I said guys, this entire document breaking down everything that we covered in depth as well as some additional things that you can read through and our complete Karpathy Obsidian Vault Starter Kit is available for free in our school community, Stride AI Academy. So make sure to check that out. It's 100% for free to join, no paywall or anything like that. And you can network with myself as well as other like-minded AI entrepreneurs, enthusiasts. We have some really cool people in there and I'm excited to continue building this community with you guys. So that's pretty much it for this video guys. I hope you guys got some value from this. I definitely tried to explain everything and be as in-depth as possible because I know on the surface just reading his tweet, it's maybe a little bit difficult to understand. So I wanted to cover everything and show you the exact process and literally just give you the starter kit so you can get going with this as soon as possible because it is very valuable when you start using Claude Code with something like Obsidian and you actually start managing your second brain in an efficient way. Now I've personally been using Obsidian probably for about the last four to five months. I have a few different vaults that I've been using and it's helped me tremendously on many different facets from content creation, business, personal life, really anything for knowledge management. And I'm excited to share more about my custom Obsidian Vaults in the future. So if you want to stay up-to-date with that as well as Claude Code tutorials, AI tutorials, make sure to like this video, comment down below your thoughts. If I missed anything or if you have any insights or ideas about this Obsidian Vault or Karpathy or really just anything about Claude Code and this video in general, let me know on down below and stay tuned for future uploads because I plan on giving you guys a immense amount of value on this channel. So if you're new here, smash that subscribe button to stay up-to-date, join the Stride AI Academy down below. And I hope you guys have an amazing week. I will see you in the next video guys. Keep hustling, keep grinding and of course guys, accelerate your stride. Take care.