llm-wiki-agent

LLM Wiki Agent

A Claude Code plugin that gives your agent persistent, structured memory. Drop PDFs, markdown files, or git repos into a folder — the agent extracts, compiles, and maintains a structured wiki. Ask domain questions and get answers grounded in your sources, with cross-references and citations.

Install

Prerequisites: Claude Code installed and working.

# 1. Register the marketplace (one-time, no manual clone needed)
claude plugin marketplace add https://github.com/gal-Tab/agent_knowledgebase

# 2. Install the plugin
claude plugin install llm-wiki-agent

# 3. In any project, initialize a knowledge base
/wiki-init

Optional dependencies (checked by /wiki-init):

• pymupdf4llm — for PDF ingestion: pip3 install --user pymupdf4llm
• repomix — for repo ingestion: npm install -g repomix (npx fallback available)

Usage

Two human actions. Everything else is automatic.

1. Drop files into raw/

cp ~/papers/interesting-paper.pdf raw/
cp ~/notes/meeting-notes.md raw/

2. Start a session and compile

claude
# Agent detects new files automatically on session start
# Say "compile" or use /wiki-compile to process them

3. Ask questions

Ask domain questions — the agent consults the wiki first, citing your sources:

"What does the paper say about scaling laws?" "Compare the architectures described in these two papers" "What do we know about OpenAI's approach to training?"

The agent reads the wiki index, identifies relevant pages, and synthesizes answers with source citations.

What You Get Per Project

After running /wiki-init, your project gets:

project/
├── raw/                       # Drop files here
│   ├── .manifest.json         # Pipeline state (auto-managed)
│   └── .extracted/            # Extracted markdown (auto-generated)
├── wiki/                      # Compiled wiki (auto-managed)
│   ├── index.md               # Content catalog
│   ├── sources/               # One summary per raw source
│   ├── entities/              # People, orgs, products, tools
│   ├── concepts/              # Ideas, frameworks, definitions
│   └── comparisons/           # Cross-source analysis
├── tools/
│   ├── extract-pdf.py         # PDF → markdown
│   └── extract-repo.sh        # Repo → markdown
├── wiki-schema.md             # Domain rules (customize this!)
└── HANDOFF.md                 # Session persistence

Each project has its own isolated knowledge base. The plugin provides the engine; your project owns the data.

How Compilation Works

Two-phase process:

1. Source page — Agent reads extracted markdown, creates wiki/sources/{slug}.md with structured summary, and lists candidate entities/concepts
2. Resolution — For each candidate, agent checks the wiki index. Existing pages get updated. New pages are created if they meet the thresholds in wiki-schema.md. Index is updated, git commits track everything.

Thresholds (configurable in wiki-schema.md):

• Entities: Created if central to a source's thesis or mentioned in 2+ sources
• Concepts: Only domain-specific or novel concepts (not common knowledge)
• Comparisons: Only on explicit request or when sources clearly contradict

Customizing for Your Domain

Edit wiki-schema.md after running /wiki-init. The most important part is the Domain Description — it tells the agent what counts as "common knowledge" vs worth a dedicated page.

Example for ML research:

## Domain Description
Machine learning research. Covers model architectures, training techniques,
scaling laws, and benchmark results.

You can also customize entity types, page sections, creation thresholds, and compilation rules.

Commands

Command	What it does
/wiki-init	Initialize a knowledge base in the current project
/wiki-compile	Compile new or updated files from raw/ into wiki pages
/learn-capture	Capture work-lessons (corrections, playbooks, insights, patterns) into the compound-learnings store

The wiki-query skill auto-invokes when you ask domain questions — no command needed.

Naming conventions

Commands, skills, and hooks follow one scheme so the surface stays predictable as it grows: