🧠 LLM Wiki Agent

An open-source CLI that turns Claude into a self-maintaining knowledge base.
Knowledge is synthesized at ingest time — not rediscovered on every query.

The Idea

Most RAG systems make the LLM do all the heavy lifting at query time — searching raw docs and synthesizing answers from scratch on every request.

LLM Wiki flips this: Claude reads and synthesizes your documents when you add them, compiling structured markdown pages. When you ask a question, Claude reads pre-built wiki pages — not raw source text.

Basic RAG   →  question → search raw docs → LLM answers from scratch (every time)
LLM Wiki    →  ingest doc → Claude compiles wiki → question → Claude reads wiki

The wiki grows smarter with every document added. Contradictions are flagged. Entities are cross-referenced. Knowledge compounds.

Architecture

llm-wiki/
├── CLAUDE.md              ← Agent instructions: how to maintain the wiki
├── agent.py               ← CLI entrypoint (interactive TUI)
│
├── raw/                   ← Your source documents (immutable, never modified)
│   └── assets/            ← Downloaded images
│
└── wiki/                  ← LLM-generated knowledge base
    ├── index.md           ← Catalog of all pages
    ├── log.md             ← Chronological action log
    ├── overview.md        ← High-level synthesis across all sources
    ├── sources/           ← One summary page per ingested document
    └── entities/          ← One page per person, concept, tool, or topic

Three Layers

Requirements

• Python 3.10+
• An Anthropic API key

Install

Clone and run from source:

git clone https://github.com/S7hrenik/llm-wiki
cd llm-wiki
pip install anthropic rich

Create a .env file in the project root with your API key:

ANTHROPIC_API_KEY=sk-ant-...

PyPI package coming soon — pip install llm-wikibase will be available on the first release.

Quick Start

python agent.py

You'll land in an interactive menu:

╔══════════════════════════════════════════════════════════════════════╗
║                                                                      ║
║   ██╗     ██╗     ███╗   ███╗    ██╗    ██╗██╗██╗  ██╗██╗          ║
║   ██║     ██║     ████╗ ████║    ██║    ██║██║██║ ██╔╝██║          ║
║   ██║     ██║     ██╔████╔██║    ██║ █╗ ██║██║█████╔╝ ██║          ║
║   ...                                                                ║
║      Knowledge synthesized at ingest time — not on every query      ║
║                                                                      ║
╚══════════════════════════════════════════════════════════════════════╝

  [1]  🚀  Init     Bootstrap a new wiki
  [2]  📥  Ingest   Add a document or URL to the wiki
  [3]  🔍  Query    Ask a question, get a cited answer
  [4]  🔧  Lint     Health-check for gaps & contradictions
  [5]  📊  Status   Wiki stats at a glance
  [0]  👋  Exit

Step-by-step

1. Init — bootstrap the wiki (run once):

Choose: 1

2. Ingest — add a document or URL:

Choose: 2
File path or URL: raw/my-paper.md

Claude creates a source summary page, extracts 5–15 entity pages, updates the index, and logs the action.

3. Query — ask anything:

Choose: 3
Your question: What are the key contributions of the Transformer architecture?

Claude reads the pre-compiled wiki and responds with citations to specific pages.

4. Lint — audit the wiki for contradictions, broken links, and gaps:

Choose: 4

5. Status — quick stats at a glance:

Choose: 5

  📄  Source pages   3
  🏷️   Entity pages   27
  🕐  Last ingest    2026-05-22 20:09
  📁  Wiki dir       /home/user/my-wiki/wiki

How It Works

Ingest Flow

When you run Ingest on a document:

1. Claude reads the full source document
2. Writes a source summary page to wiki/sources/<slug>.md
3. Creates or updates 5–15 entity pages in wiki/entities/ (people, concepts, tools, organizations)
4. Updates wiki/index.md with the new entries
5. Flags any contradictions with existing pages
6. Appends a timestamped entry to wiki/log.md

Query Flow

When you run Query:

1. Claude reads wiki/index.md to find relevant pages
2. Drills into the most relevant entity and source pages
3. Synthesizes a cited answer — every fact links back to the wiki page it came from
4. Optionally saves the answer as a new wiki page (explorations compound)

Lint

Audits the entire wiki for:

• Contradictions between pages
• Orphaned pages not referenced in index.md
• Entities mentioned but missing their own page
• Stale or incomplete pages
• Cross-reference gaps

Tech Stack

CLAUDE.md — The Schema

CLAUDE.md is the agent's operating manual. It defines:

• Folder layout and file naming conventions
• How index.md and log.md are structured
• Entity page format (front matter, sections, contradiction flags)
• Cross-referencing rules between pages
• Contradiction handling (never silently overwrite — preserve both versions with citations)
• Token budget hints for query mode (which files to read first)

You can edit CLAUDE.md to customize how the wiki behaves for your domain.

GitHub Actions

All PRs must pass CI and Lint before merging into main.

Contributing

Pull requests are welcome. For major changes, open an issue first.

git clone https://github.com/S7hrenik/llm-wiki
cd llm-wiki
pip install anthropic rich pytest ruff black
python -m pytest tests/

Make sure ruff check . and black --check . pass before submitting.
All changes go through a branch → PR → checks green → merge workflow.

See CONTRIBUTING.md for full guidelines.

License

MIT

Built with Claude by Shrenik Purvant