AI Wiki — Structured Knowledge Encyclopedia

A CLI-based knowledge wiki system that stores, searches, and manages knowledge learned by AI assistants as structured YAML data.

한국어 README

Korean User Guide | Purpose-Specific Wiki Guide

Security Notice: Local mode remains the default and must not be exposed to a network. Optional team mode adds authenticated sessions, API tokens, RBAC, CSRF, rate limiting, audit logs, and sensitive-field encryption; install it with pip install "ai-wiki[team]". Full repository encryption remains the operating system's responsibility.

---

Why AI Wiki?

AI coding agents (Claude Code, Gemini via Antigravity CLI, GPT Codex) have no built-in way to retain and reuse knowledge across sessions or projects — every conversation starts from scratch. AI Wiki is a structured knowledge store that agents read and write directly. Save a piece of knowledge once, and every agent in every project can access it automatically — no copy-paste, no repeated research.

---

Features

• AI-First Protocol — Stable JSON envelopes for deterministic agent reading and writing
• Budgeted Context — Hybrid evidence packages constrained to an agent token budget
• Evidence Citations — Document ID + JSON Pointer citations linked to source verification
• Safe Partial Writes — RFC 6902 patching with optimistic version checks and quality gates
• Autonomous Writeback — Validated creation and pending drafts without full-document replacement
• Structured YAML Storage — Accumulate knowledge as key-value structured data, not markdown prose
• FTS5 Full-Text Search — SQLite FTS5-based keyword search for Korean and English
• Vector Semantic Search — sentence-transformers + sqlite-vec embedding search
• Hybrid Search (RRF Fusion) — Combines FTS5 + vector search via RRF (Reciprocal Rank Fusion): score = Σ 1/(k+rank_i)
• Quality Gates — 5-level automatic evaluation of confidence, sources, and verification
• Auto Cross-Reference — Automatic detection and bidirectional linking of related documents
• Git Auto-Commit — Automatic Git history recording on every document change
• AI Agent Skill Integration — Auto-generates skill files for Claude Code (~/.claude/skills/), Gemini via Antigravity CLI (~/.gemini/config/skills/), and GPT Codex (~/.codex/skills/)
• Wiki Name Customization — Name entered during init is displayed in the web UI header
• Token Optimization — Efficient operation with large document sets via DB metadata queries
• Purpose-Specific Variants — Install isolated law, labor, tax, business, research, or custom wikis backed by one shared engine
• Safe Lifecycle — Backup, restore, migration, upgrade rollback, uninstall, isolation audit, and skill routing audit
• Temporal Evidence Ledger — Reconstruct current, historical, and previously known claims without overwriting history
• AI Wiki Missions — Revision-pinned plans, task leases, evidence review, pause/resume, and Codex/Gemini handoff
• Retrieval Trust Loop — Independently labeled calibration candidates with holdout gates and rollback
• Read-Only Connectors — Git, web, Google Drive, Notion, and Slack snapshots with provenance and permissions

---

Installation

User Installation (pip)

Requires Python 3.11+

pip install ai-wiki
ai-wiki init ~/my-wiki

Upgrade an existing installation:

python -m pip install --upgrade ai-wiki
ai-wiki upgrade-skill
ai-wiki reindex
ai-wiki vindex
ai-wiki doctor

AI Agent Workflow

Agents use context instead of manually chaining keyword and vector search. Every primary command returns a stable JSON envelope.

ai-wiki capabilities
ai-wiki context "How does optimistic concurrency protect wiki updates?" --max-tokens 4000 --require-vector
ai-wiki get <document-id> --fields id,title,content.facts,sources
ai-wiki record-use <context-id> --citation "doc:<id>#/content/data/facts/0" --outcome answered
ai-wiki temporal as-of <document-id> --at 2026-01-01T00:00:00Z

Reusable knowledge can be written without replacing the whole document:

ai-wiki patch <document-id> --operations-file patch.json --if-version 3 --dry-run
ai-wiki patch <document-id> --operations-file patch.json --if-version 3
ai-wiki create --document-file document.json --dry-run
ai-wiki create --document-file document.json

Source-free knowledge is stored as a low-confidence pending draft and excluded from normal context retrieval. Existing v1 and v2 files are not rewritten by reads. Modified legacy documents become v2; only documents receiving temporal data are lazily saved as v3.

Mission work uses the separate ai-wiki-missions skill. AI Wiki stores the approved plan, run, lease, evidence, review, and handoff; Codex or Gemini does the actual file, command, browser, and research work. See docs/MISSIONS.md.

Developer Installation (source)

git clone https://github.com/j-dev-team/ai-wiki.git
cd ai-wiki
python -m venv .venv
source .venv/bin/activate   # Linux/macOS
# .venv\Scripts\activate    # Windows
pip install -e ".[test]"

Windows Note: Use python instead of python3.

Purpose-Specific Wiki

ai-wiki variant install legal-team-wiki --preset law --output-dir D:\dev --agent codex --lang ko

See Purpose-Specific Wikis for lifecycle and audit commands.

Interactive Init Flow

ai-wiki init runs interactively:

1. Language Selection — English / 한국어
2. Wiki Name — Displayed in the web UI header (saved in .ai-wiki.yaml)
3. Agent Selection — Choose which AI agents you use (default: Claude Code)
4. Preset Selection — Category structure matching your wiki's purpose
5. Self-Reference Seed — Creates a detailed schema-v2 document with the wiki architecture, AI workflow, sources, verification paths, safety policy, and current engine version

Agent Selection UI:

Which AI agents do you use? (comma-separated numbers)
  1. Claude Code
  2. Gemini via Antigravity CLI
  3. GPT Codex
Select [1,2,3] (default: 1):

• Default: 1 (Claude Code only)
• Multiple: 1,2 → Claude Code + Gemini via Antigravity CLI
• All: 1,2,3
• Skills are installed only to the selected agent paths (saved in .ai-wiki.yaml)

Gemini uses Antigravity CLI (agy). Install it on Windows with irm https://antigravity.google/cli/install.ps1 | iex, then run agy once and authenticate with the Google account that owns the Gemini subscription.

Preset	Description
general	General knowledge (default)
tech	Technology & development focused
business	Business & management focused
research	Research & academic focused

---

Quick Start

# Search documents (hybrid: FTS5 + vector auto-combined)
ai-wiki search "python"

# List documents (recent first, default 50)
ai-wiki list

# Filter by category + sort by title
ai-wiki list --category technology/python --sort title

# Pagination
ai-wiki list --limit 20 --offset 40

# Get a document
ai-wiki get <document-id>

# Create a document
ai-wiki create \
  --title "Python Basics" \
  --category "technology/programming" \
  --tags "python,programming" \
  --confidence 0.9 \
  --source "https://docs.python.org" \
  --content-stdin << EOF
type: knowledge
what: Python is a general-purpose interpreted programming language
creator: Guido van Rossum
release_year: 1991
paradigm:
  - object-oriented
  - functional
  - procedural
use_cases:
  - web development
  - data science
  - automation
  - AI/ML
EOF

---

How It Works

1. You ask your AI agent a question
2. The agent automatically searches the wiki via skill trigger
3. If found   -> agent uses the knowledge to answer
4. If not found -> agent researches, then saves new knowledge to the wiki
5. Next time any agent asks the same question -> instant answer from wiki

Your Question
     |
     v
+-----------+   skill   +---------+   search   +------------------+
| AI Agent  | --------> | ai-wiki | ---------> | AI Wiki          |
|           |           |  (CLI)  |            | (YAML + DB)      |
+-----------+           +---------+            +------------------+
     ^                      |                        |
     |         hit          |<-- found --------------+
     |                      |
     |         miss         v
     |              +--------------+
     +-- answer <-- |   Research   |
          + save    | (web / docs) |
                    +--------------+

All projects and all agents share the same wiki — knowledge accumulates automatically over time.

---

Web UI

# Start web UI
ai-wiki-web

# Default port: 5000 → http://127.0.0.1:5000
# Wiki name is read from the `name` field in .ai-wiki.yaml

See the security notice at the top of this document.

---

CLI Command Reference

Basic CRUD

Command	Description
ai-wiki init [path]	Initialize a new wiki (directory structure + config)
ai-wiki create	Create a new document
ai-wiki get <ID>	Get a document by ID
ai-wiki update <ID>	Update an existing document
ai-wiki delete <ID> --confirm	Delete a document
ai-wiki list	List all documents (--sort, --category, --limit, --offset)
ai-wiki destroy [path]	Destroy a wiki (remove skill files, env vars, directory)
ai-wiki upgrade-skill	Upgrade skill files to the latest version

Search

Command	Description
ai-wiki search "query"	Hybrid search (FTS5 + vector, auto-combined via RRF)
ai-wiki vsearch "query"	Semantic vector search
ai-wiki similar <ID>	Find documents similar to a given document
ai-wiki tag <tag>	List documents with a specific tag
ai-wiki tags	List all tags and document counts

Quality Management

Command	Description
ai-wiki quality <ID>	Single document quality report
ai-wiki quality-all	Batch quality check for all documents
ai-wiki verify <ID>	Update document verification date
ai-wiki verify <ID> --human	Mark as human-verified
ai-wiki verify-queue	List documents needing verification
ai-wiki review <ID>	Generate verification checklist

Wiki Maintenance

Command	Description
ai-wiki reindex	Rebuild search index
ai-wiki vindex	Rebuild vector index
ai-wiki lint	Wiki health check
ai-wiki lint --fix	Health check + auto-fix
ai-wiki maintain	Run lint + quality + todo together
ai-wiki backlinks <ID>	List documents referencing this document
ai-wiki sync-backlinks	Bulk sync all bidirectional backlinks

Analysis & Exploration

Command	Description
ai-wiki stats	Access statistics (Top N views/searches)
ai-wiki gaps	Gap analysis by category
ai-wiki todo	Auto-collected task list for the wiki
ai-wiki stale	List outdated documents (default: 90 days)
ai-wiki discover	Find isolated, low-quality, or stale documents
ai-wiki path <ID1> <ID2>	Shortest path between two documents (BFS)
ai-wiki cluster	Document topic clustering
ai-wiki history <ID>	Git change history for a document

Export / Import

Command	Description
ai-wiki export <ID>	Export as Markdown or YAML
ai-wiki export-all	Batch export all documents
ai-wiki ingest <file>	Ingest a source file (auto-creates stub document)

---

Hybrid Search

ai-wiki search automatically combines FTS5 keyword search and vector semantic search.

• RRF (Reciprocal Rank Fusion) algorithm merges both result sets
• Returns a unified hybrid_score field
• Vector search (sentence-transformers, sqlite-vec) is included in the default installation — no extra setup needed

# Hybrid search (default)
ai-wiki search "machine learning python"

# Pure vector search only
ai-wiki vsearch "machine learning python"

---

Document Structure (YAML Schema)

schema_version: 2
id: tech-python-abc123
title: Python Basics
category: technology/programming
tags: [python, programming]
metadata:
  confidence: 0.9
  document_version: 1
  created_at: 2026-01-01T00:00:00Z
  modified_at: 2026-01-01T00:00:00Z
  verified_at: 2026-01-01T00:00:00Z
  author: ai-agent
  maturity: mature
  completeness: 0.85
sources:
  - id: src-1
    url: https://docs.python.org/3/
    title: Python documentation
    retrieved_at: 2026-01-01T00:00:00Z
relations:
  - target_id: tech-django-xyz789
    type: related_to
    direction: outgoing
    weight: 0.8
    source_ids: [src-1]
content:
  type: technology
  data:
    what: Python is a general-purpose interpreted programming language
    facts:
      - Python was created by Guido van Rossum.
      - Python was first released in 1991.
    use_cases: [web development, data science]
verification:
  - path: /content/data/facts/1
    level: verified
    source_ids: [src-1]
    verified_at: 2026-01-01T00:00:00Z
history:
  - at: 2026-01-01T00:00:00Z
    action: created
    fields: [what, facts]
    note: Document created
extensions: {}

Schema v2 validates field types, date/time zones, HTTP(S) source URLs, content types, required content fields, relation weights, and source references. Unknown top-level fields are rejected. Custom content types must be registered in .ai-wiki.yaml before use.

Existing v1 documents remain readable. Migrate them explicitly after upgrading:

ai-wiki migrate-schema            # dry-run and validation report
ai-wiki migrate-schema --apply    # backup, atomic conversion, index rebuild
ai-wiki schema-json --legacy > schema.json # bare integration JSON Schema

File Storage Path

articles/<category>/<subcategory>/<slug>.yaml

Example: articles/technology/programming/python-abc123.yaml

---

Quality System

Verification Levels

Schema v2 stores verification records outside user content and targets claims with JSON Pointer paths.

Level	Weight	Description
unverified	0.0	Not verified
sourced	0.3	Has a source
verified	0.8	Verified
corroborated	0.8	Cross-verified
disputed	0.2	Disputed
human_verified	1.0	Verified by a human

Maturity Stages

Stage	Description
stub	Minimal info, needs enrichment
draft	Basic info present
review	Under review
mature	Complete document

Quality Score Calculation

score = structure keys (15%) + word count (20%) + sources (10%) +
        tags (5%) + related docs (10%) + confidence (10%) + verification rate (30%)

---

Directory Structure

Wiki Directory (after ai-wiki init)

my-wiki/
+-- articles/           # Document YAML files (subdirectories by category)
+-- data/
|   +-- wiki.db         # Search index (SQLite FTS5)
|   +-- vectors.db      # Vector embedding index
+-- sources/            # Ingested source files
+-- logs/               # Operation logs
+-- .ai-wiki.yaml       # Wiki configuration

Source Code (for developers)

src/ai_wiki/
+-- cli.py              # CLI entry point
+-- models.py           # Article data model
+-- storage.py          # File/DB storage layer
+-- index.py            # SQLite search index
+-- vector.py           # Vector search engine
+-- quality.py          # Quality validation engine
+-- schemas.py          # Type-specific schemas
+-- catalog.py          # Catalog builder
+-- wikilog.py          # Operation logger
+-- web.py              # Web UI (Flask)

---

Wiki Operations

Single Wiki (Default)

Install one wiki and all AI agent projects on the system can automatically access it.

ai-wiki init ~/my-wiki

→ See AI Agent Skill Integration for details.

~/.claude/skills/my-wiki/   <- Claude Code
~/.gemini/config/skills/my-wiki/   <- Gemini via Antigravity CLI
~/.codex/skills/my-wiki/    <- GPT Codex

Project A --+
Project B --+-- Share a single wiki
Project C --+

Multi Wiki

Run multiple independent wikis on a single system.

Creating Wiki Instances

# Work wiki
ai-wiki init ~/work-wiki
# -> Wiki name: "Work Wiki"
# -> Preset: business

# Personal knowledge wiki
ai-wiki init ~/knowledge-wiki
# -> Wiki name: "My Knowledge"
# -> Preset: general

Switching Wikis

Switch the target wiki using the AI_WIKI_ROOT environment variable.

# Search work wiki
AI_WIKI_ROOT=~/work-wiki ai-wiki search "client"

# Search personal wiki
AI_WIKI_ROOT=~/knowledge-wiki ai-wiki search "python"

Agent Skill Integration

ai-wiki init auto-generates skill files for each selected agent, one per wiki.

~/.claude/skills/
  +-- work-wiki/SKILL.md        (AI_WIKI_ROOT=~/work-wiki)
  +-- knowledge-wiki/SKILL.md   (AI_WIKI_ROOT=~/knowledge-wiki)

~/.gemini/config/skills/  ...  ~/.codex/skills/  (same pattern)

Project A --+
Project B --+-- Access both wiki skills
Project C --+

→ See AI Agent Skill Integration for full details.

• Data: Fully isolated per wiki (separate articles/, data/)
• Program: Single shared ai-wiki CLI
• Skills: Independent skill files per wiki, accessible from all projects

---

Environment Variables

Variable	Description
AI_WIKI_ROOT	Wiki root directory path (required)

Automatically registered during ai-wiki init.

# Manual setup (Windows)
setx AI_WIKI_ROOT "C:\Users\you\my-wiki"

# Manual setup (Linux/macOS)
export AI_WIKI_ROOT="$HOME/my-wiki"

---

Development / Testing

# Activate virtual environment
source .venv/bin/activate   # Linux/macOS
.venv\Scripts\activate      # Windows

# Run tests
pytest

# Reinstall package (after source changes)
pip install -e .

Release maintainers should follow the Korean release workflow. The checked PowerShell workflow is available as scripts/release.ps1.

---

Dependencies

Package	Purpose
click	CLI framework
pyyaml	YAML parsing/serialization
flask	Web UI server
sqlite-vec	Vector embedding search
sentence-transformers	Multilingual embedding model
pecab (optional)	Korean morphological analysis

All dependencies including vector search are installed automatically via pip install ai-wiki.

---

AI Agent Skill Integration

Skill files are auto-generated during ai-wiki init. When an agent receives a knowledge-related question, the skill auto-triggers and uses the ai-wiki CLI to search, query, and create documents.

Agent	Skill Path
Claude Code	~/.claude/skills/<wiki-name>/
Gemini via Antigravity CLI	~/.gemini/config/skills/<wiki-name>/
GPT Codex	~/.codex/skills/<wiki-name>/

Skills are installed only to the agents you select during init.

• ai-wiki init — Creates skill files in the selected agent paths
• ai-wiki upgrade-skill — Updates skill files to the latest version (reads agents from .ai-wiki.yaml)
• ai-wiki destroy — Removes skill files from selected agent paths
• If agents field is missing, defaults to ["claude"] (backward compatible)

# Upgrade skill files to latest version
ai-wiki upgrade-skill

---

License

MIT License. See LICENSE for details.