twlaw

Queryable local access to Taiwan's national laws and regulations database (全國法規資料庫).

繁體中文說明：README.zh-TW.md

Why

The MOJ Open API only serves whole-database zip dumps. There is no search, no per-record lookup, and chapter headings are flattened into the article list with indentation as the only clue to structure. twlaw downloads those dumps, reconstructs the hierarchy for every article, and stores the result in a local SQLite database you can query.

Install

uv add twlaw          # or: pip install twlaw

Getting started

Building the local database is a separate, explicit step. Do it once:

from twlaw import LawDB

db = LawDB()          # opens ~/.twlaw/law.db (created if missing)
db.refresh()          # downloads all 4 datasets — ~2 min, ~520 MB on disk

After that, every query is local and fast (single-digit milliseconds). Nothing touches the network again until you call refresh() yourself.

from twlaw import LawDB

db = LawDB()                              # reopens the existing database
if db.is_empty:                           # guard for first run
    db.refresh()

for hit in db.search("扣繳義務人", limit=5):
    print(hit["law_name"], hit["article_no"])
    print("  ", hit["chapter_path"])

所得稅法 第 94 條
   第 四 章 稽徵程序 / 第 四 節 扣繳

API

LawDB(path=None)

Opens (and creates if needed) the SQLite store. Defaults to ~/.twlaw/law.db; pass a path to keep it elsewhere. Usable as a context manager.

with LawDB("./law.db") as db:
    ...

db.refresh(categories=("law", "order"), langs=("zh", "en"))

Re-downloads datasets and replaces their rows. Takes ~2 minutes for everything. Returns how many laws were stored per dataset. Narrow the scope if you only need part of it:

db.refresh(categories=("law",), langs=("zh",))   # Chinese statutes only, ~20 s

Safe to re-run — rows are replaced, not duplicated.

db.search(query, lang="zh", category=None, limit=50, include_repealed=False)

Full-text search over article text. Returns a list of dicts, best match first:

db.search("營業稅", category="law")        # statutes only, skip 命令
db.search("income tax", lang="en")        # English corpus
db.search("設籍", include_repealed=True)   # include （刪除） articles

db.get_law(law, lang="zh")

One law with all its articles, or None if not found. Takes a law name or a MOJ law code — use whichever you have. Returns the metadata fields plus an articles list of {seq, article_no, content, chapter_path}.

law = db.get_law("所得稅法")        # by name
law = db.get_law("G0340003")       # same law, by code

law["name"]              # 所得稅法
law["id"]                # G0340003
law["modified_date"]     # 20260911
len(law["articles"])     # 198

Names match exactly, not by prefix — "所得稅法" gives you 所得稅法, never 所得稅法施行細則. Use list_laws(name_like=...) or search() when you don't know the exact name.

db.get_article(law, article, lang="zh")

One article by the number you'd cite it by, or None if there is no such article. 4 is 第 4 條 and "4-1" is 第 4 條之一 — you never deal with list positions or MOJ's label strings.

db.get_article("所得稅法", 1)       # 第 1 條
db.get_article("所得稅法", "4-1")   # 第 4 條之一 — a distinct article from 第 4 條

{'seq': 8, 'article_no': '第 4-1 條', 'article_key': '4-1',
 'content': '自中華民國七十九年一月一日起，證券交易所得停止課徵所得稅…',
 'chapter_path': '第 一 章 總則 / 第 一 節 一般規定',
 'law_id': 'G0340003', 'law_name': '所得稅法'}

Every article everywhere carries this article_key, so a search() hit can be re-fetched or cited directly:

hit = db.search("扣繳義務人")[0]
db.get_article(hit["law_name"], hit["article_key"])

db.list_laws(category=None, lang="zh", name_like=None)

Law metadata without article bodies — for browsing or building a picker.

db.list_laws(name_like="所得稅")
db.list_laws(category="law")

db.update_date() / db.is_empty

MOJ's own publication date for the stored data, and whether anything is stored yet. Use update_date() to decide if a refresh() is worthwhile.

Examples

Look up a single article:

print(db.get_article("所得稅法", "4-1")["content"])

Read one law article by article:

law = db.get_law("所得稅法")

for a in law["articles"]:
    print(a["article_no"], a["chapter_path"])
    print(a["content"])

Find which laws mention a term:

for hit in db.search("營業稅", limit=10):
    print(hit["law_name"], hit["article_no"])

List every tax law, using MOJ's own classification:

for l in db.list_laws():
    if "賦稅" in l["moj_category"]:
        print(l["id"], l["name"])

Chunking for RAG

Each article already carries its chapter path, so a chunk is interpretable on its own without extra work:

law = db.get_law("所得稅法")

for a in law["articles"]:
    chunk = f"{law['name']} {a['article_no']}\n{a['chapter_path']}\n{a['content']}"
    print(chunk)

Coverage

English records are a separate, smaller corpus — not a translation of every Chinese record. A law present in both shares the same law_id, so get_law(id, lang="en") gives you the English text of the same statute.

Notes

• Concurrency. WAL mode is on, so many processes can read at once. Only refresh() writes.
• Search. FTS5 with the trigram tokenizer, which is what makes Chinese substring search work: the default unicode61 treats an unbroken run of CJK as one token, which would find 50 articles for 所得稅 instead of ~1,400. Queries under 3 characters fall back to LIKE.
• Storage. Plain SQLite tables (laws, articles), so the file is readable from any language, not only Python.

Data source and licensing

twlaw itself is MIT licensed, and the package ships no legal data. Laws and regulations are downloaded at runtime by refresh() from the MOJ Open API.

That data is published by the Ministry of Justice under the Open Government Data License v1.0 (政府資料開放授權條款-第1版), which permits free reuse, modification and redistribution — provided you attribute the source. If you ship a product built on this data, credit it, for example:

Data source: Laws & Regulations Database of the Republic of China (Taiwan), Ministry of Justice — https://law.moj.gov.tw/

Two things worth knowing if you build on this:

• The MOJ database is authoritative; this is a convenience mirror. twlaw reconstructs chapter hierarchy and splits articles, which is ordinary permitted adaptation — but the result is a parsed view. Cite the official text for anything that matters.
• Don't misrepresent the content. The license asks that the data not be altered in ways that make the displayed information contradict the original.

This section is a summary, not legal advice. Read the license if the distinction matters to you.