giki

Software-engineering approach to LLM Wikis.

giki treats knowledge like code. Raw documents (markdown, text, PDF) are compiled into structured wiki pages by an LLM, then managed through git with AI-powered PR review. Think of it as CI/CD for knowledge.

Chinese README | Design Spec

---

Why giki?

Most LLM knowledge tools either retrieve at query time (RAG) or generate content without quality control. giki takes a third path:

Compile, don't retrieve. Instead of searching through raw documents every time you ask a question, giki compiles sources into structured, interlinked wiki pages once at ingest time. The result is a navigable knowledge graph you can browse directly in Obsidian.

Review like code. Every change goes through a two-phase review pipeline: mechanical checks (dead links, schema validation) catch bugs deterministically, while an LLM reviewer evaluates semantic quality against your team's wiki rules. All of this runs as a GitHub Action on pull requests.

Git-native. Every AI-generated page is a normal git commit. You can git diff to see exactly what the LLM changed, git log to trace knowledge evolution, and git revert to undo bad edits. No proprietary database, no vendor lock-in.

---

Features

Two-phase compilation pipeline Analyze (extract candidate concepts from source chunks) → Synthesize (generate/rewrite wiki pages) → Crosslink (add [[wikilinks]] and ## Related blocks).

AI PR Review Bot Mechanical checks run first (zero false positives): dead links, frontmatter schema, index sync, unrelated edit detection. Then per-page LLM semantic review cites your wiki-rules.md rules by anchor (e.g. R-1 consistency). Verdicts: approve / comment / request-changes.

Git-native version control Each ingest produces a clean commit (ingest: observer.md — 3 of 3 pages). Branch isolation with --branch wiki/<topic>. Full diff/revert/rebase support.

Obsidian-compatible output Standard YAML frontmatter + [[wikilink]] syntax. Point Obsidian at your wiki/ directory and browse immediately.

Smart indexing index.md (categorized directory) and log.md (chronological timeline) are auto-maintained. No manual bookkeeping.

---

Architecture

graph TD
    subgraph Input
        S[Sources: md/txt/pdf]
    end

    subgraph "giki ingest"
        A[Analyze<br/>LLM: extract concepts]
        B[Synthesize<br/>LLM: generate pages]
        C[Crosslink<br/>LLM: add wikilinks]
    end

    subgraph Storage
        W[wiki/*.md<br/>flat, slug-named]
        I[index.md + log.md]
        G[git commit]
    end

    subgraph "giki review"
        M[Mechanical checks<br/>dead links, schema, index sync]
        R[Semantic review<br/>LLM per page + wiki-rules.md]
        V[Verdict: approve / comment / request-changes]
    end

    S --> A --> B --> C --> W --> G
    W --> I
    W --> M --> R --> V

---

Quick Start

Prerequisites

• Python 3.11+
• git
• An LLM API key (Anthropic or any OpenAI-compatible endpoint)

Install

pip install giki

Or from source:

git clone https://github.com/MeloMei/giki.git
cd giki
pip install -e ".[dev]"

Initialize a knowledge base

mkdir my-kb && cd my-kb
git init
giki init

This creates .giki/config.yaml, wiki-rules.md, wiki/, sources/, index.md, and log.md.

Configure your LLM

Edit .giki/config.yaml:

llm:
  compile:
    provider: claude          # or "openai"
    model: claude-sonnet-4-5-20250929
    base_url: https://api.anthropic.com
    api_key_env: ANTHROPIC_API_KEY
  review:
    provider: claude
    model: claude-sonnet-4-5-20250929
    base_url: https://api.anthropic.com
    api_key_env: ANTHROPIC_API_KEY

For Ollama or any OpenAI-compatible endpoint:

llm:
  compile:
    provider: openai
    model: llama3
    base_url: http://localhost:11434/v1
    api_key_env: OLLAMA_API_KEY

Then set your API key:

export ANTHROPIC_API_KEY=sk-ant-...

Ingest a document

cp ~/notes.md sources/
giki ingest sources/notes.md --branch wiki/my-first-ingest --yes

giki will analyze the source, propose wiki pages, generate them via LLM, add crosslinks, update index.md and log.md, and commit everything to the wiki/my-first-ingest branch.

Review changes

# Local review (HEAD vs main)
giki review

# Review a PR and post as comment
giki review --pr 42 --post

# JSON output for CI
giki review --json

---

Commands

Command	Description
giki init [--with-action]	Initialize a knowledge base. --with-action generates a GitHub Actions workflow.
giki ingest <path...> [--branch NAME] [--yes] [--dry-run] [--retry-failed]	Compile source documents into wiki pages.
giki review [--pr N] [--post] [--json] [--base BRANCH]	Two-phase review: mechanical checks + LLM semantic review.
giki config show | set <key> <value> | tips	Manage .giki/config.yaml.

Ingest flags

Flag	Description
--branch NAME	Ingest on this branch (creates if missing). Strongly recommended.
--yes	Non-interactive mode; accept all candidate pages.
--dry-run	Print candidate pages without generating them.
--retry-failed	Bypass hash check and re-run the full pipeline (recovers from transient LLM failures).

Review flags

Flag	Description
--pr N	Label the review with PR number N. Required with --post.
--post	Post the review as a PR comment via gh pr comment.
--json	Output structured JSON (for CI pipelines).
--base BRANCH	Base branch for diff comparison (default: main).

Exit codes: 0 = approve or comment, 1 = request-changes.

---

Review Pipeline

giki review
    |
    v
Phase 0  Context: load config + wiki-rules.md + determine diff range
    |
    v
Phase 1  Classify: NEW / UPDATED / DELETED / RENAMED (wiki vs index vs other)
    |
    v
Phase 2  Mechanical (no LLM):
           - Dead link detection (two-stage: filename -> alias)
           - Frontmatter schema validation
           - Slug pattern + length check
           - index.md sync (NEW pages must appear)
           - Unrelated edit ratio warning
    |
    v
Phase 3  Semantic (LLM per page):
           - Input: wiki-rules.md + before/after + mechanical findings
           - Output: findings with rule_id + severity + evidence + suggestion
           - Hand-written pages (no sources frontmatter) are skipped
    |
    v
Phase 4  Aggregate:
           - Any blocker finding -> request-changes
           - All approve -> approve
           - Otherwise -> comment
    |
    v
Phase 5  Output: markdown (default) / JSON (--json) / PR comment (--post)

wiki-rules.md

Your team's review criteria, versioned in the repo. Each rule is anchored by ## R-N:

## R-1
**consistency** -- severity: `blocker`
Facts in different pages must not contradict each other.

## R-2
**citation integrity** -- severity: `blocker`
Non-trivial claims must trace back to a source.

## R-5
**paragraph length** -- severity: `nit`
Paragraphs over ~8 sentences should be split.

The semantic reviewer cites these anchors in its findings (e.g. rule_id: R-2).

---

Configuration

.giki/config.yaml controls all behavior:

Section	Key	Default	Description
llm.compile	provider	claude	LLM provider for ingestion (claude or openai)
	model	claude-sonnet-4-5-20250929	Model name
	base_url	https://api.anthropic.com	API endpoint (override for gateways)
	api_key_env	ANTHROPIC_API_KEY	Environment variable name for API key
llm.review	(same keys)	(same defaults)	Independent config for review LLM
ingest	chunk_size	12000	Sliding window size in characters
	chunk_overlap	500	Overlap between adjacent windows
review	unrelated_edit_threshold	0.30	Warn if >30% of changes are outside wiki/
	severity_blocking	[blocker]	Which severities trigger request-changes
	pr_comment_collapse	true	Collapse nit findings in PR comments

llm.compile and llm.review are independent — use different providers/models for cross-validation.

---

GitHub Action

Generate with giki init --with-action, or create .github/workflows/giki-review.yml:

name: giki review
on:
  pull_request:
    paths: ['wiki/**', 'index.md', 'wiki-rules.md', '.giki/**']
jobs:
  review:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
        with: { fetch-depth: 0 }
      - uses: actions/setup-python@v5
        with: { python-version: '3.11' }
      - run: pip install giki
      - run: giki review --pr ${{ github.event.pull_request.number }} --post
        env:
          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
          ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}

---

Repository Layout

my-knowledge-base/
├── .giki/
│   └── config.yaml          # Model + threshold configuration
├── sources/                  # Raw documents (md, txt, pdf)
├── wiki/                     # LLM-compiled pages (flat, slug-named)
├── index.md                  # Auto-maintained categorized directory
├── log.md                    # Auto-maintained chronological timeline
├── wiki-rules.md             # Review rules (versioned with the repo)
└── .giki-state/              # SHA-256 tracking (gitignored by default)

---

Known Limitations (v0.1)

1. No PDF OCR -- scanned PDFs are rejected. Only text-based PDFs are supported.
2. No remote sources -- URLs, Notion, Confluence are not supported. Local files only.
3. No wikilink anchors -- [[page#heading]], [[^block]], and ![[embed]] are not supported.
4. Flat wiki directory -- no subdirectories in wiki/. Use tags in frontmatter for categorization.
5. Manual retry -- --retry-failed bypasses the hash check but doesn't persist per-page failure lists.
6. No token estimation -- cost control is manual. Monitor your API usage.

---

Roadmap

v0.2:

• Typed wikilinks ([[requires::X]], [[contradicts::Y]], etc.)
• giki branch / giki pr collaboration commands
• AI merge (resolve PR conflicts)

v0.3:

• Local web UI (giki serve -- D3 knowledge graph + full-text search)
• Q&A (giki chat -- BM25 retrieval + RAG)
• Cross-domain knowledge fusion
• giki lint --fix

---

Development

git clone https://github.com/MeloMei/giki.git
cd giki
python -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate
pip install -e ".[dev]"
pytest -q

See CONTRIBUTING.md for contribution guidelines.

---

License

MIT. See LICENSE.