Transform legal documents into a living knowledge graph. Feed in contracts, regulations, playbooks or guidance documents — the tool extracts legal concepts, maps their relationships, identifies knowledge gaps, and produces a full research environment: interactive graph, markdown wiki, source pages, gap-bridge prompts, project context files, incremental ontology, and prioritised to-do lists.
Inspired by Karpathy's LLM Wiki pattern and adapted for professional legal work.
Terminal preview
sources/ Output:
├── GDPR.pdf → output/knowledge_graph.html ← interactive D3.js graph
├── DPA_template.docx → output/graph_prompts.md ← ready-to-paste Claude prompts
└── DORA_guide.md → wiki/ ← one article per concept
wiki/sources/ ← one page per source document
gaps/gap_analysis.json ← structured gap analysis
todo/ ← prioritised action lists
.infranodus/ontology.json ← living memory (incremental)
CLAUDE.md ← project context for Claude Code
agents.md ← agent workflow configuration
| Phase | What happens |
|---|---|
| 1. Extract | PDF/DOCX/TXT → Claude extracts concepts, definitions, types, importance |
| 2. Connect | Cross-document connection mapping + structured gap analysis |
| 3. Source pages | One rich wiki page per ingested document (summary, takeaways, gaps) |
| 4. Wiki articles | One markdown article per concept with wikilinks |
| 5. Outputs | Graph + gap prompts + todos + ontology + CLAUDE.md + agents.md |
Self-contained HTML — open with any browser, no server needed:
- Force-directed D3.js graph with colour-coded node types
- Node size reflects importance; dashed rings mark knowledge gaps
- Click any node for definition, connections, related concepts
- Gap analysis panel with missing definitions and research questions
- Search bar, zoom/pan/drag
The graph structure is exported as ready-to-paste Claude prompts. Each prompt includes the underlying JSON graph so Claude reasons structurally — not just by retrieval. Paste directly into Claude Code terminal to bridge gaps.
One structured page per ingested document: summary, key takeaways, evidence/data, notable provisions, relevant concepts (as wikilinks), and open questions.
Incrementally updated across runs — new concepts and connections are merged,
not overwritten. Historical snapshots in .infranodus/history.jsonl.
Auto-generated on every run. Open the project in Claude Code and it immediately understands the knowledge base structure, active gaps, and available workflows.
Three files: sources to add, concepts to develop, research questions to answer — all derived from the gap analysis and updated automatically.
| Type | Description | Example |
|---|---|---|
regulation |
Laws and regulations | GDPR, DORA, EU AI Act |
principle |
Legal principles | Accountability, Proportionality |
obligation |
Legal duties | Data Breach Notification |
right |
Legal rights | Right to Erasure |
risk |
Legal risks | Unlimited Liability |
term |
Defined terms | Personal Data, Controller |
standard |
Technical/compliance standards | ISO 27001 |
entity |
Organisations/parties | Data Subject, DPA |
git clone https://github.com/marcoderoni/legal-knowledge-wiki
cd legal-knowledge-wiki
python3 -m venv venv
source venv/bin/activate # Windows: venv\Scripts\activate
pip install -r requirements.txt
cp env.example .env
# Edit .env and add: ANTHROPIC_API_KEY=your_key_here# Full pipeline — process all docs in sources/
python main.py
# Custom sources folder
python main.py --sources /path/to/legal/docs
# Fast run: graph + prompts + todos only (no wiki articles, no API calls for wiki)
python main.py --no-wiki --no-source-pages
# Regenerate graph/prompts/todos from saved extractions (no API calls at all)
python main.py --mode graph
# Regenerate wiki articles only
python main.py --mode wiki
# Don't auto-open browser
python main.py --no-browser- PDF (
.pdf) — via pdfplumber - Word (
.docx,.doc) — via python-docx - Plain text / Markdown (
.txt,.md)
Once built, open the project folder in Claude Code. CLAUDE.md gives Claude
full context. Example prompts:
"Summarise the main legal obligations in this knowledge base"
"What are the biggest gaps between GDPR and DORA coverage?"
"Use the prompt in output/graph_prompts.md to bridge the gap between X and Y"
"Draft a risk memo on data transfer based on the wiki"
"Which source documents cover controller obligations?"
legal-knowledge-wiki/
├── main.py # CLI entry point
├── CLAUDE.md # auto-generated — Claude Code context
├── agents.md # auto-generated — agent workflow config
├── requirements.txt
├── config/settings.yaml
├── extractor/document.py # PDF/DOCX/TXT extraction + chunking
├── analyzer/
│ ├── concepts.py # Claude API: concept extraction
│ └── connections.py # Claude API: cross-doc connections + gaps
├── reporter/
│ ├── wiki.py # Concept wiki article generator
│ ├── source_wiki.py # Source document page generator ← NEW
│ ├── visualizer.py # D3.js HTML graph generator
│ ├── prompt_export.py # Gap-bridge prompt exporter ← NEW
│ ├── project_docs.py # CLAUDE.md + agents.md generator ← NEW
│ ├── ontology.py # Incremental .infranodus/ manager ← NEW
│ └── todo.py # Prioritised todo generator ← NEW
├── sources/ # ← drop documents here
├── wiki/ # generated concept articles
│ └── sources/ # generated source pages
├── concepts/ # raw JSON extractions
├── connections/ # connection JSON
├── gaps/ # gap analysis JSON
├── output/ # HTML graph + gap prompts
├── todo/ # action lists
└── .infranodus/ # living ontology (incremental)
This project defaults to claude-opus-4-5 — the most capable Claude model, producing richer concept extraction, more precise relationship mapping, and deeper gap analysis. The tradeoff is speed and cost: expect ~45–60 minutes for a typical set of 5 legal documents.
Switch to claude-sonnet-4-5 for faster, cheaper runs (~5–10 minutes) with only a minor reduction in extraction quality. Recommended for iterative testing; use Opus for final knowledge base builds.
anthropic:
model: claude-opus-4-5 # or claude-sonnet-4-5 for faster/cheaper
max_tokens: 4096
extraction:
max_chunk_size: 8000
overlap: 200
graph:
max_nodes: 150
open_browser: true| Project | Description |
|---|---|
| Legal AI Toolkit | Claude Code agent + Make automation |
| Contract Scanner | Single-contract R/Y/G risk assessment |
| Contract Bulk Analyzer | Cross-portfolio analysis |
| Legal GPT Reviewer | Provider-agnostic reviewer (Claude + OpenAI/Groq) |
| Legal Knowledge Wiki | This project — knowledge graph from legal docs |
MIT © Marco De Roni 2026