Original video in Chinese.
Key Takeaway
- There are two kinds of knowledge bases, and they call for different methods. For my own use, the material must stay in my hands (Obsidian + Wiki). For others to use, I put it on a website and make it available through RAG + MCP, so members’ Agents can call on it as needed. The core principle is that accumulated knowledge should stay in your hands and remain usable across models and Agents.
- For my own knowledge base, I no longer insist on traditional RAG. With larger context windows, I first try supplying the original text directly. Retrieval is still necessary when there is too much material, or when I want to lower costs or maintain accuracy. My approach is to use Obsidian CLI to call its native precise search (keywords + structure + tags + properties), with a Wiki as a “map” that preserves the concepts, views, and relationships already organized, so the Agent does not have to start from scratch every time.
- For a knowledge base others use, I use RAG + MCP. Website content is split into chunks and given a vector index, alongside keyword matching. MCP provides search, original-text reading, and directory browsing. Content that used to be mainly for people to read can now become an external reference inside their own Agent workflows.
Back in 2024, I made a whole series of videos about building a personal knowledge base, introducing different tools and methods. We talked about vector databases, RAG, and all of that. Many of you first found me and joined the newtype community on Knowledge Planet (Zhishi Xingqiu) because you needed a knowledge base.
So, two years later, how do I use a knowledge base today? In this video, I’ll walk through the whole setup. I divide my knowledge bases into two categories: “for my own use” and “for other people to use.” Different users call for different methods.
Let’s start with the knowledge base I use myself.
For me, one thing has never changed: I need to keep the material in my own hands. Whatever model or Agent I switch to later, I should still be able to use everything I’ve accumulated, instead of starting over whenever I change tools.
Compared with two years ago, setting up a knowledge base is much easier today. I no longer build traditional RAG systems, because model context windows have grown much larger and can hold more original text at once.
But long context is not a cure-all.
When there really is a huge amount of material, exceeding the context window, or when I want to reduce the input and lower costs, I still need retrieval.
And fitting everything into the context window does not mean the model can maintain the same retrieval and reasoning accuracy across all that content.
So I did two things to help the model retrieve information and find the material more easily.
First, I use Obsidian’s native search capabilities.
One powerful Obsidian capability has long been overlooked: its precise search. It does more than match keywords. It lets you combine keywords, note structure, tags, and properties as search conditions, which makes its search highly precise.
That raises a question: how do you let an Agent call Obsidian’s native search?
It’s simple. There is an official CLI. Ask your Agent to configure it, then tell it to prioritize Obsidian CLI whenever it needs to look something up in this knowledge base.
Second, I built a Wiki.
Improving retrieval alone is not enough. If your knowledge base is a mess, the Agent will still struggle to find things. So I built a Wiki for mine.
This Wiki presents the contents of the knowledge base in a more structured way: the concepts and themes I repeatedly discuss, for example.
But it is more than a directory. It also organizes the views I’ve discussed and the relationships between them. Once something has been worked through, the Agent does not have to start from scratch every time. New content can be added later, allowing this accumulated knowledge to keep growing.
For the Agent, the Wiki is like a map. It tells the Agent what the knowledge base is for, what knowledge has already been organized, and where to find the original text when needed. This really does improve retrieval efficiency.
I implemented this approach in newtype OS quite a while ago. I’ve been using it for a long time, and it works very well. You can give it a try.
That is the knowledge base I use myself. So what do I mean by a knowledge base for other people?
On newtype.pro, I publish all my past posts, articles, video scripts, and courses. But those are for people to read. What if a user wants to give my content to an Agent as a reference?
I organized all the website content into a knowledge base. Here, I use two technologies: RAG and MCP.
First, RAG. In my system, I split the text into chunks and build a vector index. This lets a model find the closest content by meaning. Alongside semantic matching, the website also uses keyword matching to help the model find relevant material.
To let an Agent call the knowledge base, I also need a way to connect to it. I use MCP.
newtype MCP provides functions for searching content, reading original text, and browsing the directory. Once members install this MCP for their Agents, they can call those functions.
Previously, this content was mainly for you to read. Now, I want it to enter your workflow and help you get things done.
This gives members an external knowledge base they can connect to their own workflows. When your Agent is dealing with your problem, you can simply have it refer to my content as needed.
That is how I currently use knowledge bases.
For my own knowledge base, I keep the material in my hands. I have the Agent retrieve it through Obsidian, then use a Wiki to preserve knowledge that has already been organized. Even if I change models or Agents later, I can keep using all of it.
For the knowledge base other people use, the material lives together on the website side. I handle retrieval, reading, and access control there, then use MCP so members’ Agents can conveniently call on it for their own problems and workflows.
The method depends mainly on where the material lives, who uses it, and how an Agent can find and understand it. Methods can change, but your accumulated knowledge should stay in your own hands and keep being useful.
OK, that’s it for this video. If you want to understand AI, reclaim individual sovereignty, and find like-minded people, join the newtype community. See you next time!