Romanian Law Data

Zstd Parquet of the Romanian legal corpus (documents, articles, paragraphs, relationships) for DuckDB. From legislatie.just.ro (Ministry of Justice), scraped via the public HTML detail pages.

Rebuilt daily by GitHub Actions. Download from Releases.

Tables

Table	Content	Rows
documents	One row per act (LEGE, OUG, HG, ORDIN, DECIZIE, ...) with metadata + full text	~187k
articles	One row per article, parsed from documents.content	~993k
paragraphs	One row per alineat, the finest citation unit (e.g. art. 188 alin. (1))	~1.96M
relationships	One row per directed edge (act → target): repeals, amends, suspends, ...	varies

English column names (type, document_number, published_at, gazette_number, ...) with Romanian COMMENT ON metadata in create_views.sql.

Subject lenses

create_views.sql pre-filters documents per canonical code and for case law:

View	Filters
constitution	Constituția României (1991, republicată 2003)
civil_code	Legea 287/2009
penal_code	Legea 286/2009
labor_code	Legea 53/2003 (republicată)
civil_procedure_code	Legea 134/2010 (republicată)
penal_procedure_code	Legea 135/2010
tax_code	Legea 227/2015
case_law	CCR + ÎCCJ decisions

Usage

Download the latest release, then point DuckDB at the parquet files:

mkdir -p data
gh release download -R scriptogre/romanian-law-data --pattern '*.parquet' --pattern 'create_views.sql' --dir data

import duckdb
conn = duckdb.connect()
conn.execute(open("data/create_views.sql").read())
conn.execute("""
    SELECT document_citation, link, article_citation, content
    FROM articles
    WHERE document_id IN (SELECT id FROM penal_code)
      AND article_number = 188
""").fetchall()

Pipeline

etl.extract documents        sweep /Public/DetaliiDocument/{id} (resumable)
etl.extract relationships    fetch outgoing actiuni cache (resumable, sharded)
etl.transform                clean, parse, validate, write parquets, build FTS

extract documents and extract relationships checkpoint their progress into release-asset caches (documents-cache, relationships). The portal throttles per IP (~1 req/s sustained), so both run in CI as sharded matrices across many runners; each run resumes from the cache.

transform reads the document cache, runs the cleanup + parse pipeline, validates against Pandera schemas, builds the DuckDB FTS index, and writes the release parquets.

uv sync
uv run python -m etl.extract documents
uv run python -m etl.extract relationships
uv run python -m etl.transform

License

Public corpus from the Romanian Ministry of Justice. This repo provides format conversion + pipeline tooling.