📖

LLM Wiki

LLM知识维基 — AI维护的持久知识库

A knowledge base that grows smarter over time. Instead of re-deriving answers from raw documents on every query, the LLM builds and maintains a structured wiki — a persistent, compounding artifact of knowledge.

一个随时间变得更聪明的知识库。LLM不是每次查询都从原始文档重新推导答案，而是构建和维护一个结构化的维基——一个持久的、不断积累的知识产物。

Based on Karpathy's LLM Wiki pattern
The Core Insight 核心洞察

Wiki, Not RAG

维基而非RAG — 为什么持久知识优于检索增强

❌ Traditional RAG

Upload documents → chunk → embed → retrieve → re-synthesize on every query. Work is thrown away after each response.

上传文档→分块→嵌入→检索→每次查询重新合成。每次回答后工作被丢弃。
→

✔ LLM Wiki

Ingest once → build wiki pages → knowledge compounds. Each query enriches the wiki. The system gets smarter over time.

摄取一次→构建维基页面→知识积累。每次查询丰富维基。系统随时间变得更聪明。
Architecture 架构

Three Layers

三层架构 — 来源、维基、模式
📄

Layer 1: Raw Sources

Immutable curated documents — articles, papers, notes, files. Never modified after ingestion.

↓
📖

Layer 2: The Wiki

LLM-generated markdown pages — summaries, entity pages, cross-references. Updated with every ingest and valuable query.

↓
⚙

Layer 3: The Schema

CLAUDE.md — the configuration document defining wiki structure, page format, operations, and quality rules.

Operations 操作

Ingest, Query, Lint

三种核心操作
📥

Ingest 摄取

Feed a new source document. The LLM reads it, creates/updates 10-15 wiki pages, updates the index, and logs the operation.

输入新的源文档。LLM读取它，创建/更新10-15个维基页面，更新索引并记录操作。
🔍

Query 查询

Ask a question. The LLM searches wiki pages, synthesizes an answer, and saves valuable outputs as new pages.

提出问题。LLM搜索维基页面，综合答案，并将有价值的输出保存为新页面。
🔧

Lint 校验

Health check for contradictions, orphan pages, missing cross-references, and outdated information.

健康检查：矛盾、孤立页面、缺失交叉引用和过时信息。
Why It Works 为什么有效

The Maintenance Problem, Solved

维护问题，解决了

💤 Humans Get Bored

Every knowledge base, wiki, and note system eventually dies because humans abandon maintenance. The enthusiasm fades, the pages rot.

每个知识库、维基和笔记系统最终都会消亡，因为人类放弃维护。热情消退，页面腐烂。

🤖 LLMs Don't

LLMs never get bored. They maintain, cross-reference, and update tirelessly. The wiki stays alive and current with zero human effort on upkeep.

LLM永远不会厌倦。它们不知疲倦地维护、交叉引用和更新。维基保持活力和时效性，无需人类维护。

📈 Knowledge Compounds

Each ingest enriches existing pages. Each query can create new ones. The wiki grows exponentially more valuable over time — like compound interest for knowledge.

每次摄取丰富现有页面。每次查询可以创建新页面。维基随时间以指数方式变得更有价值——就像知识的复利。

🎓 Humans Focus on What Matters

Curation, analysis, and strategic thinking. Let the LLM handle the grunt work of organizing, summarizing, and cross-referencing.

策展、分析和战略思考。让LLM处理组织、总结和交叉引用的苦力活。
"The wiki is written once and then kept current, not re-derived on every query."
— Andrej Karpathy
Get Started 开始使用

Set Up Your Own LLM Wiki

搭建你自己的LLM Wiki

1. Clone the repo

Get the wiki scaffold with CLAUDE.md schema, index, and log.

获取包含CLAUDE.md模式、索引和日志的维基脚手架。
git clone https://github.com/gewenbo888/llm-wiki.git cd llm-wiki

2. Add your first source

Drop any document (article, paper, notes) into the sources/ folder.

将任何文档（文章、论文、笔记）放入sources/文件夹。
cp ~/my-article.md sources/

3. Run Claude Code to ingest

Claude reads CLAUDE.md, understands the schema, and builds wiki pages from your source.

Claude读取CLAUDE.md，理解模式，并从你的来源构建维基页面。
claude "Ingest sources/my-article.md into the wiki"

4. Query your wiki

Ask questions — the answers come from your structured wiki, not raw re-retrieval.

提出问题——答案来自你的结构化维基，而非原始重新检索。
claude "What does the wiki say about [topic]?"

5. Watch it grow

Every ingest and valuable query adds to the wiki. Over weeks and months, it becomes an invaluable knowledge asset.

每次摄取和有价值的查询都会丰富维基。经过数周和数月，它会成为无价的知识资产。