justicedao/ipfs_argentina_laws-ir

Argentina legislation IR (CID-keyed sparse GraphRAG)

Research retrieval release of endomorphosis/ipfs_argentina_laws (revision 0361102728979d41f9e2953f8b03fcc4243d3039) packaged in the skillcenter-huggingface-release/v3 thin-client layout.

Not legal advice. This is a research snapshot. The official gazette / authentic source of Argentina prevails over this corpus. Retrieved documents and graph edges are retrieval evidence only.

Primary key: entry_cid (CIDv1 raw sha2-256 of a canonical identity record). Integer document_index values are compact shard pointers, not identities.

Counts

Field	Value
Laws	8643
Articles	92392
Canonical docs	101035
BM25 terms	86072
BM25 postings	9327357
Graph nodes	109681
Graph edges	1025923
Vectors	101035 × 384-d thenlper/gte-small

Schema mapping

Source tables data/laws.parquet + data/articles.parquet. Required identifiers (laws.id, articles.id, articles.law_id) are fail-closed: the packager refuses to invent missing ids.

Source	Canonical
laws.id	source_id, law_id
laws.title	title
laws.text	body
articles.id	source_id, article_id
articles.law_id	law_id, parent_law_id
articles.title / text	title / body

Column-name drift check (Malta vs Germany): same names; article_count is int64 on Malta and int32 on Germany; Germany eli is often null.

Index layout

Zstandard parquet shards with at most 4,096 rows.

• indexes/bm25_keyword_shards.parquet maps lexical term ranges to BM25 posting shards.
• indexes/vector_chunks.parquet stores semantic routing centroids; rows inside each shard are sorted by decreasing cosine to the shard centroid.
• indexes/corpus_chunks.parquet maps document ranges to canonical corpus shards.
• data/graph/nodes and data/graph/edges are the property graph.
• data/graph/adjacency/{incoming,outgoing} stores score-ordered neighbor pages.

BM25: Okapi k1=1.2, b=0.75, title_weight=5, body_weight=1.

Graph: one node per entry_cid, facet nodes for jurisdiction / law_id / language / source, BM25_NEIGHBOR_OF edges (k=8), and article CONTAINED_IN law edges when articles exist.

Query

python -m pip install pyarrow pandas numpy huggingface_hub sentence-transformers
python scripts/query_country_laws_hf.py --local-dir . bm25 "constitution" --top-k 10
python scripts/query_country_laws_hf.py --local-dir . vector "money laundering" --top-k 10
python scripts/query_country_laws_hf.py --local-dir . graph neighbors <entry_cid>

Remote:

python scripts/query_country_laws_hf.py --repo-id justicedao/ipfs_argentina_laws-ir bm25 "constitution"

Normalize + generation code lives in country_laws_ir/ in this dataset repo.

Exclusions

ipfs_belgium_laws, ipfs_portugal_laws, and ipfs_lithuania_laws are incomplete Wayback harvests and are excluded from this indexing pipeline.

Provenance

Packaged by country-laws-ir. Upstream collector and official license remain those of endomorphosis/ipfs_argentina_laws. CIDs identify local content; they do not prove public IPFS pinning.