---
title: "Community AI Notes | 62 knowledge-base cards that give an Agent something to cite"
type: "Topics"
locale: "en"
url: "https://longbridge.com/en/topics/43458981.md"
description: "A user distilled the investing ideas in 417 community posts into 62 knowledge cards, uploaded them into a LongbridgeAI knowledge base and wired it to an Agent — which now checks those notes first rather than leaning on the model's general knowledge."
datetime: "2026-08-18T11:38:07.000Z"
locales:
  - [en](https://longbridge.com/en/topics/43458981.md)
  - [zh-CN](https://longbridge.com/zh-CN/topics/43458981.md)
  - [zh-HK](https://longbridge.com/zh-HK/topics/43458981.md)
author: "[LongbridgeAI](https://longbridge.com/en/profiles/10010333.md)"
---

# Community AI Notes | 62 knowledge-base cards that give an Agent something to cite

Yesterday, a fellow investor shared a trick for making an Agent smarter — a custom **knowledge base**.

She kept coming across cola's posts on investing philosophy in the community, and the more she read, the more she felt these ideas should "live in your head in one piece, and sit with you through every investment decision". The catch: people forget, and cola is prolific — hundreds of posts, and going back to find the right one is a project in itself.

> So she came up with a fix: **put them into LongbridgeAI's「Knowledge Base」and turn them into a dictionary the Agent can look things up in.**

Here's what she did:

1.  Had Claude Code pull cola's community posts with the Longbridge CLI — **417** in all
2.  Went through them one by one, kept the **23** that lay out reusable methodology, and turned those into **62 knowledge cards** uploaded to the knowledge base
3.  Built an Agent and connected the knowledge base in its settings — **effectively handing it a filing cabinet it can open any time**

Her own summary is refreshingly grounded: "Nowhere near cola's level, of course. But as an assistant that locates his published thinking for me whenever I need it, it does the job."

-   The Agent is already published to Explore Agents, so anyone can use it 👉 [Cola Method Mentor](https://longbridge.com/ai/agents/4lmo5k71wgr3h/chat)
-   You'll also find it in the app under「Explore Agents」👇

P.S. Publishing an Agent to Explore Agents is now open to every user,your Agent deserves to be seen by the whole community!

The original post: [I built myself a budget-version cola as an adviser](https://longbridge.com/topics/43432718)

## Some things you want AI to follow every time — and saying it once isn't enough

This isn't only this user's problem. You probably have things you **want AI to follow every time** too:

-   An investing philosophy you strongly agree with
-   Good books you've read, and the notes you wrote on them
-   Hard-won lessons from your own mistakes
-   A glossary of technical terms in both languages

Can't you just tell the AI in the chat? For this round of conversation, yes. But hundreds of posts or a whole book were never going to fit in a chat box — and a new session puts you back at square one.

> By default AI can only answer from the **「general knowledge」** it was trained on. Your own body of thinking is not something it knows.

**That is exactly what a knowledge base is for**: put the things that outlast market cycles and get reused into it, then attach it to an Agent. When a relevant question comes up, the Agent searches your personal knowledge base first and **answers from what it finds**, so the answer has something to cite.

## Building your own knowledge base is easier than it sounds

The entry point sits at the top of the LongbridgeAI Agent Platform workspace (**Agents / Knowledge Bases / Skills**) — click「Knowledge Bases」→「Create Knowledge Base」.

Four steps in all: **Create Knowledge Base → Upload File → Text Settings → Process and Complete**.

🔗 Go to [「LongbridgeAI Agent Platform」](https://longbridge.com/ai/agents)

🔗 Read the [「LongbridgeAI Agent Platform help docs」](https://longbridge.com/ai/docs/)

### Step 1「Create Knowledge Base」asks: should the Agent match words, or match meaning?

「Knowledge Base Retrieval Settings」gives you three options:

🔤 **Full Text Search ——「match the words」**

Whichever passage contains your words verbatim is the one it pulls up. Ctrl+F with muscles.

> Ask「**What does Covered Call mean**」and those words are sitting right there in the material — it hits every time.

🧠 **Vector Search ——「match the meaning」**

It isn't tied to the words you picked; it goes after what you meant, so a different phrasing still lands.

> Ask「**I'm down 20% this year and it hurts, how should I adjust**」— not one technical term in that sentence, yet it knows you're asking about mindset and position sizing, and still finds the matching material.

🎯 **Hybrid Search ——「use both」**

It runs both and then picks the best matches. **This is the recommended default**.

> In one line: **question carries proper nouns →「match the words」; question is plain speech →「match the meaning」; you get both kinds → Hybrid.**

### Before Step 2「Upload File」: let AI clean it up first

The knowledge base takes a range of file types — DOCX / PPTX / PDF / MD / TXT / CSV / XLSX / HTML / VTT / JPG / PNG, up to 5 per batch and 10MB each. For plain text that is very generous: this user's knowledge base runs to **nearly 30,000 characters** and still comes to just **68 KB**.

But being uploadable isn't the same as being retrievable. Most of what we read was written for people, and once it's cut into small chunks it may no longer explain one thing clearly — so before uploading, it's worth letting **AI pre-process** it into a shape that retrieves better.

What goes wrong varies by document. For example:

-   Upload a batch of technical docs and the file-header metadata ends up inside the first chunk
-   A table flattened into a single line, leaving a dense run of `|` and `---`

Letting AI sweep through these first avoids a fair few of them.

### Step 3「Text Settings」is the one that matters: check whether it split correctly

Once the file is up, the system **cuts it into small chunks** (1024 characters per segment by default), and every retrieval from then on works chunk by chunk.

> Chunk quality decides retrieval quality outright — this is the step worth spending time on.

You can preview the chunking before you save and process, and the test is simple:

-   **If a chunk holds exactly one complete idea, the split is right;**
-   **If a sentence is cut in half, or three unrelated topics are crammed together, the segment length needs adjusting.**

What this user did was reorganise the long posts into cards, each saying one thing — the 62 cards came out as **64 segments**, roughly one card per segment.

When the chunking looks right, hit「Save and Process」— the system runs Step 4「Process and Complete」, and the knowledge base is built.

## Once it's built, attach it to an Agent

A knowledge base does nothing on its own; it has to be attached to an Agent before it does any work. Two places to attach it:

-   **Agentic Chat**: in the「Knowledge Base」section of the Agent config page, click「+ Add Knowledge」
-   **The Agent node in Chatflow**: switch on「Knowledge Base」in the node editor panel and click `+` to choose the bases to attach (several is fine)

> ⚠️ One detail that is easy to miss: **the Agent works out「which questions belong to this base」from the base's「Name」and「Description」**, so the description has to spell out what it covers (for example「Investing knowledge base, currently including options data」).

Once that's set, whenever your question falls inside what the base covers the Agent retrieves automatically: **the reasoning trace shows a「Knowledge Base Retrieval」step, and the answer itself marks its sources with** `**Document +N**` **tags**.

## Keeping it in shape once you're using it

-   **Add content**: click「**Add file**」on the document list page to top it up any time
-   **Edit content**: open a document and you see every segment — you can **edit the text and add keywords**, and insert a custom note at the start
-   **Check how it's doing**: the list has a「**Recall count**」column — how many times this material has actually been pulled up

**「Recall count」is the knowledge base's health check**: a number stuck at 0 might mean nobody asks this kind of question, or it might mean **the content is right but retrieval can't reach it**. Run a few typical questions through「**Recall testing**」and you'll know whether to go back and fix the segments and keywords.

## You can share it with everyone, too

A finished Agent can be published to **Explore Agents** for the whole community to use!

Once it's public, the Agent still consults your knowledge base when other people talk to it.

As for whether the material you worked so hard on can be taken away — **it can't**. Even with「Allow Fork」switched on, what the other person copies is only the **Agent configuration and Skills, not your knowledge base or your keys**.

> Let others **use** your Agent → opening access is enough  
> Let others **build on top of it** → switch on「Allow Fork」  
> Don't want your orchestration exposed → leave Fork off

> 📖 Full settings (retrieval weights, Top K, Score threshold and the other advanced options) are in the official docs: [https://longbridge.com/ai/docs/basics/capabilities/knowledge-base](https://longbridge.com/ai/docs/basics/capabilities/knowledge-base)

## Join in: try it, build on it, submit your own

-   👉 **Try it & build on it**: give [Cola Method Mentor](https://longbridge.com/ai/agents/4lmo5k71wgr3h/chat) a go, and browse Explore Agents for more Agents published by Longbridge and by other users.

-   👉 **Show us what you've made**: there are more and more ways to turn an idea into something real — **Longbridge API, CLI, Skills, the Agent Platform**. One of them will fit.

**Got an idea? Build it!**

Once it's built, **@LongbridgeAI** **to submit it** — the next community showcase could be yours.

### Related Stocks

- [SNDK.US](https://longbridge.com/en/quote/SNDK.US.md)
- [TSLA.US](https://longbridge.com/en/quote/TSLA.US.md)
- [AVGO.US](https://longbridge.com/en/quote/AVGO.US.md)
- [SCCO.US](https://longbridge.com/en/quote/SCCO.US.md)
- [NVDA.US](https://longbridge.com/en/quote/NVDA.US.md)

## Comments (11)

- **奇迹的交易员cola · 2026-08-18T12:46:05.000Z · 👍 1**: Mark, please forward this.
- **涨跌平常心 · 2026-08-18T12:19:29.000Z**: After reading, I feel the real barrier isn't in the features, but in the organization: filtering 417 posts down to 23, then breaking them into 62 cards
  - **好运绵绵兔** (2026-08-18T12:21:43.000Z): Actually, AI can handle a large portion of the work; I just do a final review of its output (I also force myself to read through it once, otherwise it really doesn't sink in 😂)
- **卖飞专业 · 2026-08-18T12:14:09.000Z**: Are all plaza benefits now unlocked? Previously, you had to pay for a package to upload.
- **亏钱不缺席 · 2026-08-18T12:13:08.000Z**: Saved, going to open a position this week
- **亏钱天花板 · 2026-08-18T12:10:13.000Z**: The agent relies on the knowledge base's name and description to determine which queries should be routed to it. In other words, if the description is vague, the agent won't use the KB regardless of how well it's built.
- **吉祥满盈 · 2026-08-18T12:05:32.000Z**: I didn't realize the knowledge base was for this; I always thought it was meant for enterprise use.
- **好运绵绵兔 · 2026-08-18T11:53:20.000Z · 👍 5**: Thanks to the official account for reposting my post!After everyone uploads their Agents to the plaza, you can also check the usage stats ( ╹▽╹ )This private advisor generated 82 conversations yesterday—a pretty solid result.
- **贝塔狗BD · 2026-08-18T11:47:29.000Z · 👍 1**: Tested it, and it really analyzes based on the approach cola mentioned earlier! It offers a deeper industry perspective.
  - **贝塔狗BD** (2026-08-18T11:48:00.000Z): Can be my advisor