Transcription
Last week, Andrej Karpathy, a prominent AI researcher, released a super useful framework called LLM Wiki that uses LLM to supercharge your research. It's a really good framework. However, there's a critical piece missing. It's not aware of itself.
The problem that Karpati himself pointed out is that when you interact with your standard AI and you upload some files and you want to get a result, what happens is that it retrieves uh some chunks of text using rag retrieval augmented generation and gives you the result and then every time you interact with this system, it's always going to do the same thing and forget what it's done in the past. So you don't have a knowledge base that is evolving with you.
Andrej Karpathy proposed to resolve this problem by building so-called LLM wiki which you can open in any IDE or Obsidian that basically retrieves all your raw papers or notes organizes them in a folder which we call raw folder here and then it generates a wiki where it lists all the concepts and the connections between them the main ideas. So it creates an organization for your knowledge and then you can use this structure to then interact with it and to generate insight.
The problem with the system however is that even though you have the structure you still need something to navigate through it because if you just ask your LLM to produce the results it's going to extract the concepts that you have listed the questions the data but then it's going to produce the most probable outcome because that's what LLMs are trained to do. This is where knowledge graphs can be very useful.
For the last 10 years, I've been working on InfraNodus, a knowledge graph visualization tool that represents ideas as network and then applies graph science metrics to identify the main ideas and clusters of them inside, but most importantly identify the gaps in your knowledge which you can then use to generate new ideas. So, I'm going to show you how you can augment this LLM wiki with knowledge graphs so that you can generate much more pertinent insights and so that you can direct your LLM's thinking through that structure that it created. I'll just show you a couple of things it can do and then I'm going to demonstrate how you can set the whole system up from zero.
So, let me show you here. I have an already prepared wiki that I made from my documents on a research topic I'm working on. And uh what I did first is that I ingested all my papers into the raw folder here uh which you can see some papers here, some notes that I have on different topics and so on. And then I ran the skill that I'm going to also make available to you for free uh through my GitHub repo. I'm going to leave the link in the description to the video below that basically ingests all this raw material and then generates structured representation of the main concepts in those research papers, how they're connected to one another. So the main connections, some sources that it uses so that you can always get back to those sources or add more interesting research papers and then systems that it's talking about.
Okay. So normally what happens here is that once you have that wiki created here which is a structured representation of your knowledge, you could launch claude code inside that folder and then start asking questions about it. But when you ask questions, what it's going to do is that it's going to scan through your wiki, extract some concepts, and then produce the most plausible response that kind of connects everything together. It can be good for some use cases, but when you want to generate new ideas, is not so interesting because the results are going to be quite generic.
What I need first, and this is where knowledge graphs are very useful, is to get an understanding of the structure of this wiki. And so for this you can actually use knowledge graph of InfraNodus which you can use both from your favorite IDE like in this case I use cursor maybe you're using anti-gravity you can install the extension of InfraNodus here and then when you have that folder of concepts for example where I have the most important concepts listed that were extracted from the research papers I right click on it then I click InfraNodus visualize as a graph and then the extension which is opened in the side bar here is going to visualize to me what are the main topics in my research and this is very useful because first of all it reminds me what are the main ideas inside network theory you know then phase dynamics fractal structure but also shows me the smaller clusters that I want to develop in order to make this whole discourse much more coherent so I will focus on the smaller ideas like power mechanisms here for example and explore them further or frequency modulation and so on right so this helps me see what are the topics that I need to develop further. But then I can also go into gaps and ask it to show me some clusters that could be better connected and then focus on the connections between those clusters to generate new ideas.
You can also use this plug-in inside Obsidian itself. So here for example I have Obsidian and in Obsidian we have the InfraNodus plugin. So then I would open that same folder concepts and visualize it as an InfraNodus graph and then it would show me the connections between the ideas that exist in this folder give me the same topical overview and also I could identify the gaps and then by bridging the gaps and using AI to generate research questions I can then generate new ideas that bridge those different clusters together.
So this would be one approach that you can use it visually like this or you can also go back to cloud and actually use the InfraNodus mcp server which is basically a connector for your llm to think in terms of these graphs so that it would create that same graph representation of your content. So I could ask okay use InfraNodus generate graph tool to analyze the concepts in the wiki concepts folder. Okay. So then it would connect to the InfraNodus MCP tool which is basically an like an API that connects uh this whole content to Infronotus and then it extracts the list of the concepts that we have listed here as markdown files. Then you see it uses the InfraNodus generate knowledge graph tools and it's going to give me a similar insight to what I visually see. So here again I can generate an overview and then I can use InfraNodus to identify the main clusters and then ask InfraNodus to generate insights based on the gaps between these clusters. So that would be one approach to use this. You see it runs the tool then it generates the analysis and then it will generate a response. As you can see here it provides me information about the graph structure main topical clusters fractal dynamics critical transitions process coupling and also information about the gaps.
So you can use the graphs if you want or you can also use the MCP server inside cloud code itself. However, an even more advanced approach is to add knowledge graphs into the structure of this LLM wiki so that instead of just analyzing the folders with the external tool, you actually integrate knowledge graphs into the process of structuring your knowledge. So that skill that I made available actually allows you to do that. And what it does is that once your wiki is created and you generate some ideas from it, then it will create a folder InfraNodus where it's going to save so-called ontology graphs that I was talking about before on my channel which describe the main concepts and relations between them. So then later if you say want to analyze what's happening in this repository in this wiki you can go back to Obsidian for instance and then you can open that folder InfraNodus and for example if you want to see a graph of all the concepts you have here then you just click on this and Obsidian InfraNodus plug-in shows you what are the main concepts you have here and also which of them are disconnected. So that can also give you information how you can connect them better. So for example, you can see like okay, you have synchronization patterns here, but they're not so connected to, for example, fractal networks. So you select these two and then you generate an interesting question that can link them together. And this will give you an idea for how you can link disconnected clusters of your knowledge and make your whole research much more coherent and interconnected which can be very useful for improving its structure.
Okay, so this would be another approach how you could use it. You generate InfraNodus ontology graphs for everything that it does. So it's some kind of living memory of what's happening in the system and then every time you come back to it claude or whatever system you use will have access to this system of rules of of how the concepts connect in your wiki and then it will be able to extract them and generate some interesting ideas from them. So that would be another deeper approach where you integrate knowledge graphs into the actual LLM wiki.
Now that I showed you how all of this looks like, let me show you how you would actually set something like this up. So first I would create a new folder on my computer and I would give it the name of the research I'm working on. So for example, I want to make another one on finance. I click create and then I go inside and I simply open the empty folder in my favorite programming development environment which is cursor in this way. All right. So now it's completely empty. Then what I do, I open terminal inside here. And by the way, you could also do it directly through terminal. But what I like is that in something like Corsor or if you like anti-gravity, you can use anti-gravity. I can actually see the file structure on the left. That helps. Then I open cloud and because I already installed that LLM wiki skill that I'm going to link to in the description to the video below, I can then um ask it okay use LLM wiki skill to create a repository of knowledge on finance that I have. And you can even make a general query like that and see how it does. Sometimes you will need to tell it where it should take the data. Let's see if the skill will actually pick it up because I structured it in a way where it will ask you questions for how to proceed. Here it's asking us shall I use the skill lm wiki we say yes. Okay let's use the skill and then see what it does.
So first of all it's asking me okay you want to explore finance that this wiki is for just explain to me what you want to work on personal investing microeconomics and market personal finance trading and strategy. So I would say that I would like to work on personal investing. I think that will be interesting for me. And also trading and strategy. Okay. So then let's click next. And you see everything that it's going to do. It's going to ask me a question about it. And then it tells me, okay, so what kind of sources you want to feed? Articles? Yes. Books and papers? Videos and podcasts? No. But personal notes and data? Yes. Okay. And then it's telling me, okay, so what's your end goal for this finance week? So I think I want to track and learn. I want to support my decisions. And I want to also have a living reference. Submit. You see, it provided me um with some questions. I answered them. Now I click submit answers. It's going to now use the skill to structure my knowledge in the best way possible. So now while all this is happening, it's basically building something from it. And here it's telling me, okay, so how many sources you want to have? Um let's say 20 to 100 medium. Do we need entity pages for individual stocks, funds, or companies or people? Yes, let's do that. Uh, do you track your own portfolio data? Occasionally, and we submit the answers. So, you see, every time it's going to build something, this skill, it's going to ask you a few questions so it can structure its work properly just for your particular use case.
So, here it asks us a few more questions about what kind of structure we want and what kind of trading strategies we want to have. We say that we want to provide some sources to it. And as you can see, it generated a really nice structure where we have the raw folder where we will ingest all the relevant articles, notes, data and books. Then we have the folder where this knowledge will be structured. That's what I showed you at the beginning with the concepts, connections, data, entries and so on. And then uh it will have a folder with the output where it will generate interesting ideas and even a to-do folder so that you can keep track of the things that you need to do to this repository. And because it's using infronotus under the hood then it's going to also have a hidden folder InfraNodus where it's going to store knowledge graphs that will keep record of all the important connections and concepts that you have in this repository of knowledge. Now, as you can see, it also generated the cloud MD file and agents MD file that's for Codex with a description of what this project is about and what we want to achieve with it. So, that will really help the agents later to generate some interesting ideas from it. And also it added some technical files to make sure that we can uh track the changes that we do to this and uh save intermediary versions as well.
Now, we say okay, we're ready to ingest a source. So let's say that we have a source already. So I'm going to say yes. Let's let's do the source. And then it's going to tell me okay. So where do you want to take the source? And I can tell it okay I have all my research papers on the Dropbox research papers folder. Find everything relevant to finance and economics there. And you see the claude is very clever in this case because it's basically going to search that folder and extract the most interesting papers and images and other assets I have on the topic of finance, money, and economics. As you can see, it says it found a lot of material. Now, it's going to sort it into tiers of relevance and organize it. So, here it's asking us what it should actually extract. Core finance, finance and market dynamics or everything potentially interesting. I think maybe not everything but let's say that we want finance and market dynamics around 30 papers. What clo is going to do now is going to extract the content from those papers extract the content from PDFs and transform it into markdown files which is very good because then you would be able to open it using any text editor and also they would be accessible to your LLM because it will basically be like plain text. Now it's copying the files to my raw folder that all the books are copied here. And then like it says here, it's going to create wiki pages for these papers in parallel. If we go into wiki, we see that now it has the folder with sources and it's adding paper by paper into that folder. And here we have a summary of the paper. We have some key takeaways from each paper, some data and evidence and also relevance, quotes and links to important concepts and ideas that are present in this paper. So it's creating an interconnected structure of the main ideas that relate to this project which is really great and it would take you hours or even days to do that manually. Our end goal here is to extract some interesting insight. So this is why having this structure is going to be very useful.
What happened now here is that from these papers that it found were relevant, it extracted a list of concepts and also a list of connections and uh that already creates a really nice structure for our LLMs to use inside cloud in order to generate some interesting insights for us. While it's doing this job, we can actually use the InfraNodus extension that I have installed into my cursor in order to see the content of every page that we have generated here. So for example, here I have a page on balance of payments. If I want to see what it's about, I can rightclick then click in front of this visualize as a graph and then it will generate for me a really nice visual overview of the main ideas in this summary of the concept of balance of pay payments. It's about capital dynamics, exchange flow, investment trends, market insight, and so on. If I open the bond one, then if I right click on that, I can then visualize the main concepts there. So, as you can see, you're actually creating really nice visual representation of the main ideas. And if I want to see how all of my concepts connect, while of course, Claude will create some connections for me. If I don't want to rely on AI to do that, I can right click on the whole folder, then click visualize as a graph. And what's going to happen here is that InfraNodus will visualize uh all the concepts from all these different files and show me what are the most important topics. So I can see there's a lot of on bond returns and financial flows, but for example, very little on equity analysis or regression analysis. And if I decide that maybe regression analysis is quite important because it will provide me a framework through which I can analyze financial data then it's going to be quite important and I should develop this cluster more. So you see already just from this visualization that is happening live as LLM is building that knowledge base I can already optimize it by understanding which topics I should develop more what kind of sources I should seek to make that more coherent and interconnected. And if I go into the gaps analysis here, I can also see which topics are not so well connected. And then using this AI advice button, I can generate a question which is like an advanced prompt that I can feed back to my claude by selecting it and pasting it here. That will enable clot to understand a little bit better how it can develop ideas further. So, it's pretty interesting because not only you get a really nice visual overview of what's happening in your folders as LLM is building this wiki, but you can already optimize it by identifying the clusters that are underdeveloped and that you want to develop further or finding the gaps that you want to bridge with new ideas.
Actually, let me show you how that would work in practice. So here I have a gap identified in the folder on concepts and if notice shows me that there is a cluster on financial flows and on regression analysis there is a gap between them. So what I can do is I can click on the question here and then it's going to find me the statements from the folder of concepts that could be better connected and here it provides these two statements to me. But then there is also the InfraNodus log which is basically a prompt that uses the underlying graph that I can use to generate some interesting insight. So I'm going to copy that prompt with the underlying graph structure that shows the two clusters and the possible connections between them. Then I'm going to go to terminal and open another terminal window. So the second one launch clo there and say okay based on what you see in the concepts folder try to think of the connection between these clusters and then I paste this structure which I copied from the InfraNodus log with the underlying graph structure that represents these clusters. Then I provide it here and I can also use this as inspiration. These are the actual extracts from the concept documents that uh contain uh the most dense representation of those clusters. And then I can also say to claude uh also take this into account. Okay. Then I provide the statements and then you see now it's going to basically use this structural information. So I point LLM's attention to the gap that exists. I provided the underlying structure. I give it some context and then I say okay use the documents in the concepts folder in order to generate an interesting idea that would connect these ideas together. How does regression analysis connect to financial flow? And you see it's doing the work. It's extracting some relevant concepts from our wiki from the concepts themselves analyzing all these documents taking this graph structure that uh we we identified into account and then now it's going to generate an insight for us. Let's see what it generates.
So as you can see here based on this gap after doing some research it identifies a potential research question how the critical transitions and systemic fragility affect the bop a fix models reliability and it tells us do you want me to create a connection page or question page to capture any of these insights. Yes. Let's add the question and add it into to-do list because we want to research it further later. Right. So now it will use that question that it extracted that found the gap identified the gap generated an interesting research question and then it's going to add that question into our to-dos list so that we can later explore it either manually or automatically using AI and kind of like focus on something which is not yet explored in the literature but that touches upon the topics that the literature is talking about. So we know that we will have uh good quality source material to work with but we will actually come up with something new and that can be interesting for us because we will link these ideas in new ways. So here it added this into to-do and if we look at the wiki we now see that we have the concepts we have connections between some of the concepts. We have also important entities and we have some research questions that we are adding that will enable us to then guide our research and explore the connections between the topics that are not connected yet. So then we will be sure to generate new insights this way.
So this is how the system works in a nutshell. I know it's a lot to ingest, but what I recommend you to do is just to install the skill and to try it for one of the projects you have so you can test run it yourself with your own data and see what kind of insights it can generate for you. And if you see that you want this model to work differently, you can always modify that skill and add some changes into it that will suit your workflow better. And once you have something like this generated, you can interact with it using cloud code using your favorite IDE or you can even go back inside Obsidian and you will have that structure presented here. So you can use the more traditional Obsidian structure to navigate through this data, visualize the connections between different ideas and see how you can link them together in new ways. So you have Obsidian as a viewer of this information including the InfraNodus graph view plug-in or you have a more low-level representation here where you use claude or your favorite AI in order to interact with this data and to generate new ideas from it and you have the graph available here which you can use to also visualize this knowledge to find gaps inside and to see what the main concepts are and how they relate to one another and finally you can use the InfraNodus MCP server inside cloud code itself in order to generate knowledge graphs and to save them into the InfraNodus folder so that you have an ongoing memory of the conversations you're having.
So here if I want to create such graph I can go back to the first discussion we had where it ingested 20 papers on the topic. It says the wiki is ready to use open the finance folder as of old and then when you start quering it ingesting new sources or gap analysis if you like. So for example now I can say okay now that this whole wiki is ready run gap analysis on it and then look what happens here it's basically going to do the same thing that I do in the graph when I choose gaps but now it's going to read through the main concepts that it ingested and then it's going to launch the InfraNodus gap analysis tool in order to generate uh the clusters that could be better connected and then help us generate new ideas from them. So you see here it says let me run the InfraNodus analysis directly on the wiki content to identify structural gaps. So it first gathers all the wiki content from all the files that it saved into the wiki here. It will then create a query for them. Now as you can see it created the full wiki text and then it's going to use the InfraNodus generate content gaps tool via the MCP server that is connected to my cloud code and then it's going to generate content gaps based on this content and deliver me similar insights that I can get also through this visual interface either inside the cursor or anti-gravity or inside Obsidian as well if I'm using the AI questions here that identifies the gaps and then generates questions based on those gaps. But here the difference is that I don't actually need to use the graph. So if I feel uncomfortable using the graph interface, I can just delegate all this responsibility to my LLM to the MCP server and then I'll just have the readym made insights. But I find it's important to also demonstrate to you how this whole thing works even if it makes it a bit more complicated because you know what's happening under the hood that it's not going to be some hallucination but we're actually performing structural analysis here where we identify the main topical clusters find the gaps then feed those gaps to the model with the underlying graph structure and steer its attention to the parts that should be better connected so that the quality of insights we get are much less generic and much more original.
So as you can see here, it provided me the final outcome of the research it made, identified some of the priorities, added them into to-do list. So that's the stuff that I should be working on to explore these ideas further. And then it even created an ontology, an InfraNodus graph with all the main ideas inside that show how the different concepts that I have in this whole wiki relate to one another in relation to this particular exploration. So that next time when I will be talking to this folder through clot code or through any other tool uh LLM will have access not only to the whole wiki structure but also to all the relations and ideas inside and it will have a much more precise understanding of how all of these concepts relate and also what questions I should be working on to have this knowledge evolve further.
So let me know how this works for you. If you have any questions, please ask them in the comments below the video. And I hope you find this demo useful and I encourage you to try it on your own notes, ideas, research papers, books. I think you will get really interesting results. I did. Thank you.