我用DeepSeek V4搭知识库，成本降了90%，效果不输Claude Opus (English)

Generated: 2026-06-24 18:29:22

---

I Spent 28 Days Building an AI Butler That Can "Eat" My Entire Knowledge Base

Last month, I was about to blow a fuse.

Guess what? I had three devices lined up on my desk, five browser tabs open—all just to find a technical proposal I'd written the week before. I clearly remembered saving it in some folder, but no matter which keywords I tried, it just wouldn't show up.

That afternoon, I cursed at my computer screen for a solid ten minutes.

Then I slammed the table—screw it, I'll do it myself!

---

Now that I mention it, have you ever experienced this kind of despair? The more notes you write, the harder they are to find. I tried every tool on the market, one after another.

Notion? Data stored in the cloud—I could never quite feel at ease. What if one day I just can't access it?

Evernote? Its AI capabilities are basically zero. You ask it to do something, and it has no clue what you mean.

ChatGPT? You have to upload files every single time, then wait for it to finish reading, and even then, it might not read them correctly. The efficiency is so low it makes you want to smash your computer.

Dedicated AI knowledge base products? Thousands of dollars a year! On my little shoestring operation, I'm pinching pennies just for the servers—there's no way I could afford that kind of luxury.

So what happened next? I saw DeepSeek V4's API pricing—pennies per million tokens! And its performance is about on par with Claude Opus.

Right then and there, I slammed the table again: This is it!

After a week of building it, and another three days of testing, tuning, and falling into every pitfall imaginable, the system has been running on my laptop for over half a month now. The results are way better than I ever expected.

Today, I'm going to lay it all out for you—the entire setup process, real test data, and every single lesson I learned the hard way, paid for in cold, hard cash.

---

Why These Three Things Together?

My criteria for choosing tools are just three:

My data must stay in my hands. The AI must be smart enough. The price must be down-to-earth.

See? Pretty simple demands. But as simple as they are, no product on the market could meet them.

Until I combined these three—

Obsidian satisfies the first requirement. All files are .md format, stored on your computer. No internet? Still accessible. The plugin ecosystem is comprehensive. In short, it's a super-powered notepad, but its ceiling is far higher than you can imagine.

Claude Code is a command-line tool from Anthropic. What's its killer feature? It can directly operate on your local files! It's not you chatting with it and hoping it gets the gist; it actually reads and writes to your file system. That's a game-changer.

DeepSeek V4 provides the brains. Its long-context comprehension is comparable to Claude Opus 4.6 in my tests, and the price? Only a tenth of Claude's!

When I was testing it, I ran hundreds of tasks, and the credits I burned didn't even cost as much as a cup of milk tea.

String these three together, and what do you get?

An AI administrator that never sleeps, is always on call, and is incredibly cheap.

---

The Setup Process: Four Steps, and I've Already Filled in the Potholes for You

Step One: Install Obsidian

Go to the official website and download the installer for your system—available for both Mac and Windows, and it's free.

Double-click, install, done.

You've got this one.

Step Two: Install the Claudian Plugin

What's this thing? It's the bridge that embeds Claude Code into Obsidian. Skip this step, and you'd have to switch back and forth between the terminal and the editor—a terrible experience, almost as bad as not setting it up at all.

Here's how: Open Obsidian Settings, go to Community Plugins, and turn off Safe Mode. Search for BRAT in the community plugins list, install, and enable it.

BRAT is a tool that can pull beta plugins from GitHub. Claudian hasn't been listed on the official marketplace yet, so you need BRAT to get it installed.

You know, I got stuck here for half an hour. The BRAT interface is a bit confusing at first glance—you don't know where to start.

When adding a beta plugin, enter this address: github.com/YishenTu/claudian, and select the latest version. After adding it, go back to the plugin list, and the toggle for Claudian will appear.

Turn it on.

Step Three: Install Claude Code, Then Swap Out Its Backend Model

This step requires a little bit of technical savvy, but if you know how to copy and paste, you're good.

First, install Node.js. For Mac users, run:

brew install node

For Windows users, go to nodejs.org and download the LTS version. After installation, verify it:

node --version
npm --version

See the version numbers? Good, let's move on.

Now install Claude Code:

npm install -g @anthropic-ai/claude-code

Here's the crucial part.

Claude Code by default calls Anthropic's own API—expensive and not always reliable what we need to do is point it to DeepSeek.

Create a new folder called .claude in your user directory. Inside it, create a file named settings.json and fill it with this:

{
 "env": {
 "ANTHROPIC_AUTH_TOKEN": "Your DeepSeek API Key",
 "ANTHROPIC_BASE_URL": "https://api.deepseek.com/anthropic",
 "ANTHROPIC_MODEL": "deepseek-v4-pro",
 "API_TIMEOUT_MS": "3000000",
 "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1"
 }
}

Go to the DeepSeek website to buy an API Key—it's just a few bucks—and paste it in.

Then run claude in your terminal. If it shows that the model has been switched to deepseek-v4-pro—congratulations, you've configured it successfully!

Step Four: Back in Obsidian, Connect Claudian to Claude Code

In Obsidian Settings, find the Claudian plugin page and configure the environment variables. Fill in the same content from your settings.json.

Select the model as DeepSeek V4, and set the maximum context tokens as you see fit. I set mine to 32,000, which is more than enough for daily work.

Here's a particularly sneaky pitfall: Obsidian must have its command line interface enabled. You can find it in Settings > About > Advanced. Turn it on.

Otherwise, Claudian won't be able to connect to Claude Code, and all your hard work will be for nothing.

---

Real Test Results: True Gold Fears No Fire

After setting everything up, I immediately started intense testing.

Let's start with basic retrieval.

I deliberately asked a very narrow question from my library of over a thousand notes: "The on-device model deployment solution mentioned at the 2024 Shenzhen GTC Conference."

In the past? I would have to first recall roughly when I wrote it, which folder it was in, and then manually dig through layer by layer.

Now? I just typed it in.

Five seconds. Just five seconds.

The AI listed three related notes, each marked with