YouTube2Text

Карпатый Wiki Вместо RAG — Полный Obsidian Сетап Для Новичка — Transcript

by ИИшенка | AI Automation · 4,015 words · 265 segments · language en · Watch on YouTube

Full transcript

  1. 0:00Friends, so, all the hype surrounding Andrey Karpatov's article about how he's building
  2. 0:04his own Wiki based on LLM is gradually dying down. That is, everyone has moved on. But,
  3. 0:09as often happens, when the hype dies down, what remains are purely practical cases. There are many of them,
  4. 0:15they're top-notch, they're truly useful, and they can be used in practice. And today, in this video,
  5. 0:20we'll look at these cases, build such a system manually from scratch, and answer the
  6. 0:25following questions: can this system be used not just for fun and toying around,
  7. 0:30but to build truly working systems, for example, for your own company, for your own
  8. 0:36research, or for other truly practical use cases? So, the idea is this: the way neural networks currently
  9. 0:41work with our documents—most often, rag approaches, or retrieval augmented generation—
  10. 0:48is that we have a set of documents, which we first divide into
  11. 0:55a certain number of chunks and vectorize these chunks. And every time we ask a
  12. 1:00question to the neural network, it uses similarity search to find the necessary relevant chunks, returns them
  13. 1:08to itself as context, and generates a response based on this context, which works quite well.
  14. 1:14Andrey Karpaty proposes an alternative approach called Wiki LLM. He argues that
  15. 1:20we can work with complex document structures differently, preserving the full context.
  16. 1:27Our input may be the same documents as in case number one. But we empower our neural network
  17. 1:34to analyze these documents individually and find important semantic points in each
  18. 1:40one. Once these points are found, the neural network begins to connect the different semantic points
  19. 1:46. These points are converted into text documents. And these text documents now contain
  20. 1:51cross-references, that is, cross-links that link different ideas.
  21. 1:56Information about which documents are already contained in our system is stored in a special file called "index."
  22. 2:01So, it's indexing specific files and their
  23. 2:05very brief descriptions. And this leads to magical results. When we ask Lampa something,
  24. 2:10it starts by examining the index file to understand
  25. 2:14its current knowledge. Using this file, it begins searching for relevant texts that will help
  26. 2:19it answer our questions. But since these texts and files are linked to other files,
  27. 2:25which we've already taken care of, Lampa can recursively search through all the relevant files,
  28. 2:30gather the full context, and return an answer based on this context,
  29. 2:35which is gathered from various files. This is much more accurate than the standard cancer approach. Of course,
  30. 2:40this has its downsides, which we'll also discuss now. Friends, before
  31. 2:45we start building our system and achieving such amazing results and
  32. 2:51accelerating them with Nerona, I have three basic requests for you. Be sure to subscribe
  33. 2:55to the channel right now, leave a comment, and like this video so that
  34. 3:00as many people as possible can see it. Let's move on. Oh, and don't forget to drop
  35. 3:05by our pro group, take the master classes, and develop your skills in artificial intelligence. There's
  36. 3:09a ton of interesting stuff there. Let's move on. For example, could wikis like these be used to
  37. 3:14integrate work chats, work knowledge, or to immerse ourselves in
  38. 3:21resource data and then search for intersections and important insights within that knowledge? Looking ahead,
  39. 3:27I'll say that, of course, it's possible. It's a truly powerful tool. The concept of what
  40. 3:32Andrey Karpaty is talking about didn't come out of nowhere; it's based on a hypothetical system
  41. 3:38described by Weiner Busch in his essay "As We May Think." It's called Memix. It was
  42. 3:45published in 1945; by the way, I'll post a link to it in my free Telegram group. So, back
  43. 3:51then, Bush described it as an electromechanical device that allows for the creation of an autonomous
  44. 3:57knowledge base, complete with associative links and notes, which could, in turn,
  45. 4:03be transferred to other similar knowledge bases. It's all very interesting to read, but
  46. 4:09the real problem Push faced at the time was precisely keeping this base up to date.
  47. 4:16And now the development of artificial intelligence is greatly boosting these concepts, because we're
  48. 4:21shifting the conceptualization and maintenance of this database to artificial intelligence
  49. 4:27—that is, creating new connections, clearing out old, unnecessary connections, updating all
  50. 4:35indexed files—while leaving the processing of this knowledge, acquiring it, finding new
  51. 4:41connections, and thinking to humans. So, what is Andrey Karpaty writing about? Everyone
  52. 4:46reads that tweet from Andrey Karpaty, which has already garnered 20 million views, but in fact,
  53. 4:51he describes the concept of LLM wki. Let's translate it into Russian. It explains
  54. 4:56the main ideas from which one can derive the architecture of what he proposes in order to
  55. 5:01use and adapt it to truly superior quality. Incidentally,
  56. 5:06one of the cases he highlights is a business team. He says
  57. 5:10you can build internal eyelids, updated with messages, say, your internal correspondence,
  58. 5:15project documentation, calls from clients, and so on, so on,
  59. 5:20so on. We need to understand the architecture. This is the simplest part, because so many
  60. 5:25people have talked about it. It consists of three basic things. These are just three directories. The first
  61. 5:29directory is called RO, that is, some raw incoming data. The second directory is called wik,
  62. 5:36that is, Markdown or text files that your neural network itself creates, meaning you don't
  63. 5:41mess with them. And the third, probably the most important part, is the "with whom" schema. This is essentially the brain of your
  64. 5:48system, all the instructions. If you use Clcode or Codex, then this is the equivalent of your
  65. 5:54ClMD or AgentsMD file. In other words, these are the instructions that the neural network receives every time
  66. 6:03you communicate with it, which it perceives as absolute truth or rules. So, if we
  67. 6:09imagine that we're receiving a document or plain text as input,
  68. 6:14it first goes into the PO folder in the raw data. So, your basic task is to put it there. Then,
  69. 6:21your lab, any one you use, looks at the schema file, reads all the usage rules
  70. 6:28for how we process new files in R. It processes this file and, in accordance with
  71. 6:34these rules, enriches or modifies the WiK folder. In other words, it either adds new
  72. 6:40text files or updates old files. This is done to extract all the key
  73. 6:47component points from this file, establish connections between them, and if any old files
  74. 6:54or pieces of knowledge are already there, connections are also established between them and the new pieces of knowledge. And
  75. 7:01this operation we just discussed is called injest, or loading and adding new
  76. 7:08data into your system. There's a second operation called a query, or QU. QU is
  77. 7:14something different. Qury no longer works with the RA folder. Qury works with a "wik" folder, which already
  78. 7:21contains our text files and an MD index. The MD index is essentially an index file that contains a list
  79. 7:29of all the text files that the lab creates when it parses all our incoming
  80. 7:36data. This is essentially a huge collection of contents that our lab reads every time
  81. 7:42we ask it something. There's another operation called Lint.
  82. 7:47It's essentially a cleaning operation. This term comes from programming. Essentially, it periodically
  83. 7:53runs through the files we have and cleans them of unnecessary junk. Let's say
  84. 7:59there might be some duplicates, or there might be files that don't have these links between them. Then
  85. 8:04this entire system will work to either create links, delete unnecessary files,
  86. 8:10or merge several files together. In other words, it's a kind of constant self-cleaning. And again, according to
  87. 8:16Andrey Karpatov's method, this whole thing replaces the system's rag, but only on certain
  88. 8:23small volumes. Because as long as the system can place an index MD
  89. 8:28file, which contains a list of all the files, their descriptions, and other metadata, into its context, Nerronka works
  90. 8:34much more accurately, allowing us to use it and search for insights in the data
  91. 8:39we've placed there. Essentially, Andrey says this works surprisingly well if
  92. 8:44you have about 100 sources, which are essentially converted into several hundred wiki pages.
  93. 8:51which are constantly updated and brought into line with reality. This doesn't mean that
  94. 8:57if you have more than 100 sources, such a system won't work for you. But it essentially means
  95. 9:02that for each aspect of your research or your workflows, it makes
  96. 9:07sense to create a separate Wiki, which will have isolated contexts, its own
  97. 9:12index files, its own Wiki pages, and in which you can sort raw incoming
  98. 9:17files based on their type, remembering to delete, for example, outdated files, merge
  99. 9:23several files into one, and so on. Well, I suspect this sounds pretty scary,
  100. 9:28but it works great, and it's very easy to set up. Let's set up such a system together. So,
  101. 9:34to set up such a system, we'll need Obsidian. Karpaty himself uses
  102. 9:39Obsidian, and he uses this analogy. Imagine Obsidian as your IDE,
  103. 9:45that is, a kind of development environment. The LMA itself, the neural network that works with files
  104. 9:51within Obsidian, is essentially, well, let's say, a developer. And the wiki, the entire system of files
  105. 9:58created using the neural network, is essentially your code, written within
  106. 10:06your IDE. And the coolest thing is that in this system, you essentially do exactly
  107. 10:12what a human is expected to do. First, you decide which sources and which data are fed
  108. 10:18into your Obsidian, your IDE. After that, you communicate
  109. 10:24with the neural network using simple commands: "Process new files" or "Answer my question." You never actually create
  110. 10:31wiki files; that's all handled by the neural network. Your job is to actively monitor what
  111. 10:36the neural network is doing, communicate with it, and catch it when it's doing something wrong. Essentially,
  112. 10:41you do exactly what a human is expected to do when working with artificial intelligence.
  113. 10:46Since we've already discussed the fact that it makes sense to create a separate
  114. 10:51Wiki for each topic, for each project, so to speak, let's create a new directory in
  115. 10:56Obsidian called Project One. By the way, if you don't know what Obsidian is,
  116. 11:01it's a great tool for managing all your text notes
  117. 11:08and creating cross-references between them, that is, relationships between text notes.
  118. 11:13It allows you to visualize this scheme, so go to the obsidian.m website and download it
  119. 11:19for your Obsidian operating system. It's completely free. You don't need
  120. 11:23to use the paid version with cloud sync. Download it. After that, you'll see an interface
  121. 11:28similar to the one we use to create a new project. We'll also create two folders and two directories.
  122. 11:34One is called RO—this is the directory where all our source,
  123. 11:40raw data, raw files, will be stored . Let's create another directory called Wiki. This is where
  124. 11:46all the files our neural network will generate from raw files will go. According to
  125. 11:51what? According to a schema. So, we'll need a file. Let's
  126. 11:58use ClДMD. This clД MD will become the source instructions for essentially any of our
  127. 12:04clodcode requests. The big question is, what do we put in this MD? In this MD, we put a schema, a workflow that
  128. 12:12will explain to Neronki what to do with the data we put here, depending on our request
  129. 12:18. Essentially, we need to describe to Neronki everything we read in Andrey Karpatov's Wiki so
  130. 12:25that it knows how to handle loading, requesting, linting, or cleaning operations. You
  131. 12:30can give me this entire text of your Neronka and say, "Create me a schema file so
  132. 12:35you know what to do with new files or how to answer my questions
  133. 12:40when I ask them, and Neronka will create a set of rules for you, which you can then put
  134. 12:46into your ClD MD file." That's exactly what I did. I created a ClD MD file based on what's specified
  135. 12:53in Andrey Karpatov's Wiki. Essentially, all the nuances are described here so that my
  136. 13:00Neronka knows how to work with my Wiki. The structure is described here, that is, the R folder itself,
  137. 13:05what it is, what to do with it. The Wiki folder that is here, the Wiki index MD, which will be created.
  138. 13:10Automatically, Wiki Log MD, which will also be created automatically. Logging, by the way,
  139. 13:15is a very important file in this system, because Nerona will record all the raw materials
  140. 13:21it has processed there, essentially maintaining a chronology of what has already been done to
  141. 13:25avoid duplication later. And as you can see, everything
  142. 13:29that corresponds to our scheme is clearly defined here. The injest workflow is precisely the workflow for adding new files.
  143. 13:34Nerona will know that you need to first look at the rotalog, then parse it all,
  144. 13:40read it, create a summary in the Wiki folder, and create subpages for each topic it
  145. 13:46deems necessary. Well, everything we discussed, and finally, create a new entry in our
  146. 13:51log. Next, we specify the page format that suits us, in which we must specify
  147. 13:56that we would like to have links to other pages there and that Nerona should handle this for us.
  148. 14:02Then come the rules. Next comes the workflow, when we ask for answers to our questions. That is, Nerona
  149. 14:07should always start by reading the indexMD file, then process everything according to the plan, followed by linting and
  150. 14:12periodic cleaning. This can be automated, or it can simply happen once every so often, as
  151. 14:17Nerona says: lint, clean our system. And then these are very important rules,
  152. 14:23because as your system starts to grow, you'll notice all sorts of artifacts. Depending on
  153. 14:28which Nerona you're using, you'll clearly see where something is going
  154. 14:33wrong. That's why this section is so important. Here you can add new things,
  155. 14:40new instructions for Nerona, to avoid certain negative consequences. You should
  156. 14:45understand that this file is very general, designed to handle essentially any case,
  157. 14:51any ourch, or topic we're working on. But you can make any changes to this file
  158. 14:58to suit your specific needs, so it can handle your specific tasks with the required
  159. 15:03level of accuracy. I'll attach this ClD MD file, which you can copy and paste into your obsidian,
  160. 15:09to my free Telegram group. The link will be in the description. Head over there,
  161. 15:14download it, play the deck, and follow my steps. And if you're just starting out in the field of artificial
  162. 15:19intelligence, be sure to drop by our pro-group. There are great master classes on ClD,
  163. 15:25Antigravity, N8N, and a huge amount of pro-materials and Pro-Chat. The link to the pro-group
  164. 15:31is also in our description, so be sure to drop by there and level up your AI skills
  165. 15:37. We're continuing. And essentially, this is already enough to start
  166. 15:42testing our system, because we have the schematic, the catalogs, and now we can give it the first
  167. 15:49raw data. To make it easier for us to work with this entire system, I suggest you click this
  168. 15:54gear and select "Community Plugins." And be sure to install the plugin
  169. 16:00called "Terminal." This allows you to conveniently open an integrated
  170. 16:04terminal directly within your Obsidian. Here we have an "Open Terminal" button. I
  171. 16:09click this button and then "Integrated Terminal." And now my terminal has opened right in
  172. 16:14my local directory, where I'm currently located. Let's jump to the Project directory itself. That's it,
  173. 16:20we see that we're now in the Project directory. And let's call our
  174. 16:24CLOAD within this directory. Here we have our CLOAD code, ready to begin processing
  175. 16:29our requests according to what's written in the CLOAD MD directory. Let's put our
  176. 16:36first data in the folder. You should understand that you can put anything here. You can put PDFs,
  177. 16:42Word files, or just text files, because CLOAD natively handles these formats. It
  178. 16:49has built-in tools for working with them. If you're using a different neural network,
  179. 16:53all modern neural networks that work within the Command Line Interface can essentially also
  180. 17:00work with any file type. I'm creating a new file, untitled, right inside the r folder. I
  181. 17:06thought it would be cool to consider this case using a real internal company chat as an example,
  182. 17:11say, which you'll use as raw data for your process. For example, I have
  183. 17:17We've had a simplified chat in our Slack for a week now. Accordingly, there are several participants.
  184. 17:24Every day we have discussions about AI, cancer evolution, new models, how
  185. 17:31we automate certain processes, open source versus APIs. So, these are very, very different
  186. 17:37conversations. I realize there's quite a bit of gold here, for example, and I'd like to have some insights
  187. 17:43constantly sorted and saved to my internal wiki so
  188. 17:50I can then develop ideas based on what we discuss with colleagues in
  189. 17:55our work. So now I can return to my CLD code and issue just one command.
  190. 18:01If you've already forgotten, we have a CLD MD file that describes what our neural network should do
  191. 18:06depending on what we ask it to do. Accordingly, as you may recall, one of the workflows
  192. 18:12supported or defined in our MD file is inest, or workflow insertion. So
  193. 18:19, let's return to our a clD code. So, how do we say "inest new data?"? We send a request to our clD code.
  194. 18:26And we understand that now, look, our neural network knows what "injest" means. It's already found
  195. 18:33the new file and is starting to process everything we have in our directory. R. So, look,
  196. 18:38processing our RO files is finished. And the system tells us that I found the source. That's right.
  197. 18:45This is the chat for the past week. And I've highlighted five interesting topics that are being discussed here.
  198. 18:52Here we discuss graphrack. Here we have the LLM gateway, or controlling the budget
  199. 18:57that goes to neural networks. Security AI, because one of our bots was hacked. This is how
  200. 19:04agent Warflows work, especially in combination with Jira, Varnish, and Confluence, and other very
  201. 19:10interesting, cool topics that would be a shame to lose by simply leaving them in the chat.
  202. 19:16And the neural network is instructed to do this. Before making any real changes to our wiki,
  203. 19:22it needs to ask questions, that is, bounce stories off us, like, "I'm doing this,
  204. 19:27what do you think?" So, in reality, the most important thing is happening right now, and it depends entirely
  205. 19:33on us. We need to carefully review what it's found and perhaps
  206. 19:37make some adjustments. So, we need to be quite involved at this stage, because after this stage,
  207. 19:43we'll have new data entered into the wiki. And after that, let's say, in some cases,
  208. 19:47it'll be recorded. Oh, by the way, Nerona here, well, she replies in English. You can, of course,
  209. 19:52instruct her directly in the rules here that she'll only respond in Russian,
  210. 19:57because we have, ah, the entire clд MD er in English. Accordingly, Nerona herself decides
  211. 20:02in which language to respond to you. But here I'm saying, "Ahad," like, "Let's go, don't stop." As we can see,
  212. 20:08Neronka is currently working on creating eight Wiki pages for us and immediately updating the index md
  213. 20:15and l ind for our project. So, as we can see, our Neronka has finished, and it has created a whole
  214. 20:22bunch of files, including index md and log md. As we can see, our Wiki folder has now been enriched with all
  215. 20:30these files. Here is our index md. Note that it lists the sources
  216. 20:37used, the concepts (that is, all the concepts that were found), and all
  217. 20:42the relationships it deemed important and extracted from our discussion. There is a log here,
  218. 20:49specifically what was done and when. Here, it was done today. And then we created
  219. 20:54files for each topic it deemed necessary. So, here we have an LLM gateway
  220. 20:59or cost control, that is, all the nuances we have here. In other words, it very systematically and fully
  221. 21:05described what we discussed this week. How can we use this now? Let's
  222. 21:10clear our context. So, we can start working with our neural network in our session again
  223. 21:15. Now I can ask something I've been waiting for. I write that this is a quy request.
  224. 21:21What are we going to do about controlling neural network spending? I send this request to the neural network.
  225. 21:26And it should understand that this is now a different type of request, that it now needs to study the Wiki index,
  226. 21:32which it does, paying attention to what's there, building all the connections
  227. 21:38, and starting to respond to us based on these connections. The entire problem and the team's solution.
  228. 21:44Our marketing blew our entire Open AI budget in one day; I'm generating SEO articles en masse through
  229. 21:49Unlimited internal API. The LLM gateway solution uses semantic routing, meaning simple
  230. 21:55tasks will automatically be routed to Laama. 3.8 billion company GPU parameters,
  231. 22:01zero cloud usage. Only complex queries go to the cloud. Then there's the user limit,
  232. 22:07meaning Light LLM tracks costs per user for high-volume departments like
  233. 22:12marketing, along with stricter limits and scanning. This means we block sensitive
  234. 22:17data from being sent to the cloud. And the most interesting thing is, as our data grows, if we add a new chat every week,
  235. 22:23all our correspondence, to LLM. It becomes structured, it identifies all the connections, and each
  236. 22:29time our dialogue becomes more and more meaningful, more and more accurate. And the neural network
  237. 22:34, since it begins responding based on reading the initial LLM index,
  238. 22:40always has complete context about what's happening with us and the company,
  239. 22:44based on our chat. When our data volume grows, we can always run a
  240. 22:50lint command, which initiates a self-cleaning of the system. And it knows what this command means,
  241. 22:56reads the files to double-check them for the audit and give us an easy summary of
  242. 23:02what can be optimized now. So, our audit is complete. The result is that we
  243. 23:07haven't found any inconsistencies yet. Well, our Wiki is quite small, but when it starts growing,
  244. 23:12the inconsistencies are quite large. But look, we have several concepts, but
  245. 23:19there are no pages for them. That is, for example, for QDRH—this is our main vector database. It is mentioned,
  246. 23:26but there is no page for it. And the recommendation is to create this page, missing cross-
  247. 23:32references, well, and other shortcomings. And, accordingly, it asks us what it wants me
  248. 23:37to fix all this or something specific. So we tell it: Fix all. Let it fix everything
  249. 23:43it found. So, our changes are ready. Here are the problems. Here's what's done. We created
  250. 23:49two pages. We added an internal link. We've added notes and acronym expansions
  251. 23:55that weren't previously expanded. Here we see our new pages, which are now also
  252. 24:00included in our index. Look, we see that our index, which previously didn't have
  253. 24:05a DRAN file, now has this file with an internal link. We can always jump to it and
  254. 24:10view it. And now the sweetest and most magical thing is, when we switch to Graph View mode,
  255. 24:18we can observe how our relationships and points of tension, or
  256. 24:25knowledge concentrations, begin to emerge. We see how nodes are being identified that we need to pay attention to
  257. 24:30in our future work. So, we've just built a small initial version of this
  258. 24:35system. And according to Andrey, it's very cool to use in several contexts.
  259. 24:43That is, it can be used to build personal knowledge bases, for research,
  260. 24:49obviously, for reading books, so that all concepts and all intersections of relationships can
  261. 24:55be tracked this way, then accelerated along with Nerona and more firmly assimilated. This is really
  262. 25:01cool to use for your business and your team. This is exactly the case we
  263. 25:05looked at today for competitive analysis. In general, in all projects where overlap is important,
  264. 25:12it's crucial to have a neural network with rich context and quick access to it. That
  265. 25:17concludes the video. Be sure to like, comment, subscribe, and see you soon.

About this transcript

This page contains the full transcript of Карпатый Wiki Вместо RAG — Полный Obsidian Сетап Для Новичка by ИИшенка | AI Automation, generated from the public captions YouTube serves with the video. The transcript has 4,015 words across 265 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.