Original video in Chinese.
Key Takeaway
- My AI note-taking system is divided into two parts: external information processing (Anything LLM) and note content generation (Obsidian).
- Anything LLM supports multiple large models and vector databases, and can handle PDFs and WeChat Official Account articles for digesting and storing materials.
- Obsidian is my ultimate note-taking choice because it is fast, keeps data local, and has rich AI plugins such as Copilot.
- The note system filters external information through Anything LLM, turns valuable parts into Obsidian notes, and then uses AI to assist with content generation.
- For note classification, I use four categories: PROJECTS, AREAS, FLEETING, and PERMANENT, to keep everything organized.
- I emphasize that tools are secondary; the core is clear needs and logic, and building a system through workflows and tools.
Recently, I made a major upgrade to my note-taking system, adding an LLM-driven knowledge base and making some major adjustments to the overall logic.
I’m going to share my thinking and the specific approach here. You can copy the homework first, then keep adjusting it as you use it.
The whole system is divided into two parts:
The first part is processing external information.
Every day we see a huge amount of content: papers and research reports in PDF format, articles on web pages, and so on. The notes we make all start from reading and digesting this external information.
So how do you digest, store, and retrieve so much material? That’s the difficult part of this stage, and also where AI can play the biggest role.
The second part is generating note content.
There are two core questions here:
-
What kind of logic is most reasonable for classification? I used to find this really annoying: either the categories were too broad and felt meaningless, or a new note would suddenly appear and I’d find that it didn’t fit anywhere, which was just absurd.
-
Which software is the most suitable? It needs to be fast, privacy-safe, and also have AI features as assistance.
Let me start with the first part. For processing external information, the tool I use is Anything LLM.
I’ve recommended a lot of tools like this in videos and in Knowledge Planet. After using them all, Anything LLM is the one that best matches my needs. There are two reasons:
First, it can connect to mainstream LLMs, embedding models, and vector databases on the market.
For example, on the closed-source LLM side, the top three—OpenAI, Anthropic, and Google—you can use them just by entering the API key.
On the open-source LLM side, it supports both Ollama and LM Studio. Just fill in the link and port, and you’re good to go.
In a recent version update, Anything LLM also integrated Ollama directly. It provides a series of models; for example, the one I use most often is Mistral 7B, which you can download directly through the software and start using.
Some models are just too large and definitely can’t run locally, so then I’ll spend some money, run them in the cloud, and connect to them locally.
So what’s the point of having so many access methods?
I mainly use two computers in daily life:
At home, I use a desktop, the PC whose configuration I introduced before. Its performance is still okay, so it has no problem running local LLMs.
When I’m out, I bring a MacBook Pro. This machine is already very old; I bought it in 2017, and there’s no way it can run a large model now, so I can only call OpenAI’s models through the API.
Besides being able to choose different LLMs according to different computer setups, Anything LLM also lets different Workspaces use different models. For example, if the materials in a Workspace are all in English, I just use Mistral; if there’s both Chinese and English, I use qwen-1.5.
Second, in addition to supporting documents like PDFs, it can also pull content from WeChat Official Account articles.
A large portion of the Chinese information I receive comes from WeChat Official Account articles.
Tencent should have anti-scraping measures. I’ve tried many products of this type, and not every one of them can pull the content of a WeChat Official Account article through its link.
That’s my approach to processing external information. I use the AI knowledge base to store materials and help me digest them quickly. Later, when I need to find something, I can also search quickly. If this part is handled well, then the note-taking part later becomes very easy.
I’d consider myself a long-time note app user. I started with Evernote, so it’s been, what, ten years? After using so many products, my ultimate choice right now is: Obsidian.
I know someone will definitely ask: why not use Notion, which is super popular right now? Two reasons.
First, it’s too slow.
In Notion, many actions have a little bit of loading time, and I just can’t accept that. I think a note app should be like a physical notebook: open it and you can see it, flip to wherever you want.
Obsidian doesn’t have this problem. It’s especially smooth.
Second, the data is stored in someone else’s place.
In the external information processing stage above, I didn’t choose a local database because those documents and web pages are all public information, with no privacy or security issues, so I don’t mind putting them in a cloud database.
But notes are different. These are truly private data, and I absolutely will not put them into someone else’s database. This is something I’m going to accumulate for many years. If Notion ever has a problem one day, that would be troublesome.
Every note in Obsidian is an md-format file stored locally. If you want, you can copy them elsewhere at any time.
As for Notion’s AI capabilities, Obsidian has them too. This software supports community plugins, which can add all kinds of functions on top of the core, including calling LLMs.
The Copilot plugin is especially useful. You can use closed-source LLMs like OpenAI and Google, or connect to Ollama and LM Studio to use open-source LLMs.
What’s even more powerful is that it also comes with RAG capabilities built in, which can turn all your notes into a knowledge base. For example, if I ask the AI a question, it will refer to all my notes to give an answer, and the end of the answer includes sources. Clicking them jumps directly to the corresponding note.
In this way, a pipeline is formed:
First, I store all external information in Anything LLM and use AI to digest and organize it.
Then, the valuable parts of that information get turned into notes and placed in Obsidian.
Finally, when I want to write an article, AI can extract relevant content from hundreds or even thousands of notes, organize the logic, and then give it to me.
That’s the operating logic of this entire AI note-taking system.
To make note classification more organized, I watched a lot of advice from big names on YouTube, and in the end I went with this structure, divided into four categories:
First, PROJECTS, which are various specific projects. For example, my video script creation is a project. So I put all kinds of inspiration and topic ideas into it.
Second, AREAS, which are different fields. For example, AI is one field I pay attention to, and console games are also a field I pay attention to—notes from different fields naturally need to be separated.
Third, FLEETING, which is for various temporary thoughts. There’s no need to create one category after another for them; just put them all in one place.
Fourth, PERMANENT, which is the final finished product. For example, every script for each of my videos goes into this category. Because the video has already been published, the script is naturally Permanent too.
These four PAFP categories can classify and hold all the notes I want to keep, so things won’t get messy when I need to find them.
That’s my entire AI note-taking system. Its starting point is my thinking and my logic; when these logic patterns are projected outward, they become a workflow; and in that workflow, you need all kinds of tools, which is why there is the software I just introduced; workflow plus tools equals a system, one that fully expresses my thinking and my logic.
So tools are secondary. What your needs are, and how exactly you want to do things, are the most important. Everyone must not become obsessed with tools.
OK, that’s it for this episode. It’s all drawn from my own experience, and I hope it’s helpful. If you have anything you want to ask me, come find me on Knowledge Planet. See you next time!