your-ai-lawyer

Official country laws sit in git submodules. Claude Code reads them, writes a summary + wikilink graph into knowledge/, and a local ChromaDB index lets you search or ask questions with citations back to the real article. No API key. Nothing leaves your machine.

Spain is wired up (12,000+ laws from the BOE). 27 more countries are one git submodule add away.

What you get

• Run /compile es-ct --limit 10 and Claude reads ten Cataluña laws, writes summaries, extracts concepts, and wires up wikilinks.
• Output is a valid Obsidian vault. Open it in Obsidian and the graph is there.
• Two ChromaDB collections: one for the compiled summaries, one for raw article chunks. Queries hit both.
• /qa answers in prose with [BOE-A-XXXX-XXXXX, Art. N] citations. You can click through to the original.
• Recompiles skip files that haven't changed (md5 hash check). Every run is logged.
• /lint flags orphans, broken wikilinks, and untracked notes.

Quick Start

Install

git clone --recurse-submodules https://github.com/arnabdeypolimi/your-ai-lawyer.git
cd your-ai-lawyer
uv sync

Requires uv and Claude Code.

Open in Claude Code

claude .

Pick a language (once)

/setup            # interactive
/setup es         # direct

Built-in: en, es, ca, fr, it, de, pt. Any other BCP-47 code works too, it just goes straight to the model.

The choice lives in .claude/settings.json as env.OUTPUT_LANGUAGE. New /compile runs and /qa answers pick it up immediately. Existing notes stay in whatever language they were compiled in, which is deliberate — change the setting, run /compile <jurisdiction> --force, and you'll rewrite them.

Compile some laws

Start small. Cataluña with a limit of 10 is a good smoke test:

/compile es-ct --limit 10

Then scale up. Constitutional-only is ~50 files; all of Spain is 12k+ and you'll want to run it in batches of a few hundred:

/compile es --rank constitucion
/compile es

Build the search index

/index
/index --country es --compiled-only

Ask it something

/qa What are the housing rights of tenants in Spain?
/qa Does Cataluña have its own data protection laws?
/search right to education --country es --n 10

Sanity check

/lint
/lint --jurisdiction es-ct
/lint --broken-links

Slash Commands

Spain's regional codes: es for national, plus es-ct, es-md, es-an, es-pv, es-ga, es-vc, es-ib, es-ar, es-cn, es-cl, es-cm, es-cb, es-as, es-ri, es-nc, es-mc for the 17 autonomous communities.

Architecture

your-ai-lawyer/
├── legalize-es/              # Spain raw laws (git submodule, 12K+ files)
├── knowledge/                # Compiled Obsidian vault
│   ├── laws/es/              # One .md per law: summary, provisions, wikilinks
│   ├── concepts/             # Concept index files with backlinks
│   └── jurisdictions/        # Per-country overview notes
├── data/
│   ├── index.json            # Per-law compilation status
│   ├── compile.log           # Run history (NDJSON)
│   ├── manifest.json         # MD5 hashes for change detection
│   └── chroma/               # ChromaDB vector store (gitignored)
└── src/
    ├── compiler/             # parser, extractor, batch, tracker, lint
    ├── indexer/              # ChromaDB embedding pipeline
    └── query/                # Semantic search

Compilation pipeline

legalize-<country>/
  raw markdown + YAML frontmatter
          │
          ▼  /compile (Claude Code reads + writes directly)
          │   list_files.py  →  lists files needing compilation
          │   tracker.py     →  records result in index.json + compile.log
          │
knowledge/laws/<country>/
  compiled notes with [[wikilinks]]
          │
          ▼  /index (local ONNX embeddings — no API key)
          │
data/chroma/
  compiled collection + raw_chunks collection
          │
          ▼  /qa or /search
  citations: [BOE-A-XXXX-XXXXX, Art. N]

What a compiled note looks like

Every knowledge/laws/es/<identifier>.md starts with YAML frontmatter:

---
identifier: BOE-A-1978-31229
title: "Constitución Española"
country: es
jurisdiction: es
rank: constitucion
status: in_force
compiled_at: 2026-04-21
---

Then a summary, key provisions (bullets), cross-references as wikilinks, any supersedes/implements relationships, concept tags, and a path back to the raw source. knowledge/ is a valid Obsidian vault — point Obsidian at it and the graph is there.

Tracking

Three files in data/ track compilation state:

Check status directly:

uv run python -m src.compiler.tracker status
uv run python -m src.compiler.tracker status --jurisdiction es-ct
uv run python -m src.compiler.tracker log --n 20

Adding a New Country

All the raw legislation comes from legalize-dev — one repo per country, Markdown + YAML frontmatter, maintained independently.

Replace XX with a country code below and run:

git submodule add https://github.com/legalize-dev/legalize-XX.git legalize-XX
git submodule update --init legalize-XX
claude .
/compile XX --limit 20
/index --country XX
/qa What are the employment rights in France?

For the big ones (US has 60k+ sections, Portugal 109k+ norms), don't try to compile everything at once. Do it in batches by legal rank, and keep an eye on the tracker:

/compile us --rank statute --limit 50
/compile pt --rank lei --limit 50
uv run python -m src.compiler.tracker status --jurisdiction us

Available Countries

28 countries. Pick one, submodule it, compile. Spain is the only one wired up in this repo so far — everything else is a git submodule add away.

✅ = already added as submodule · ⚠️ = under active development, structure may change

Contributing

Issue first, PR second. Most useful contributions right now: adding a new country submodule, tightening the compile prompts, improving the linter, or writing a better extractor for a specific legal rank.

Acknowledgements

None of this works without legalize-dev. They scrape, clean, and version the raw government publications into Markdown; this repo just compiles a graph on top. Each country submodule is maintained over there under its own license.

Terms of Use

Personal and non-commercial use only.

This is not legal advice. I'm not a lawyer. The compiled summaries and Q&A answers are generated by a language model from publicly available government sources, which can be outdated or incomplete, and the model can be wrong. If you're making a decision that matters, talk to an actual lawyer and verify against the original source the citations point to.

By using this you accept that:

• You won't use it commercially without sorting out the upstream licensing yourself
• The maintainers aren't liable for anything that goes wrong
• Laws change. Compiled notes are a snapshot, not the current state

License

MIT for the code. Law texts in the submodules are public domain (official government publications) — check each submodule for specifics.