justicedao/patent-legal-ir-graphrag

Patent Legal IR GraphRAG

Single-dataset retrieval release at justicedao/patent-legal-ir-graphrag.

Layout matches Publicus IR GraphRAG releases (Publicus/cvefixes-security-ir-graphrag, Publicus/skillcenter-ir):

Family	Path	Notes
Corpus	data/corpus/*.parquet	dense document_index, CID keys
BM25 documents	data/bm25/documents/*.parquet	lengths + entry CID
BM25 postings	data/bm25/postings/*.parquet	sorted terms, FTS5 IDF, sparse lists
Vectors	data/vectors/*.parquet	256-d embeddings
Graph nodes	data/graph/nodes/*.parquet	structural + BM25 term vocabulary
Graph edges	data/graph/edges/*.parquet	authority edges + contains_term
Adjacency	data/graph/adjacency/{incoming,outgoing}/*.parquet	paged pointers
Routing indexes	indexes/*.parquet	remote shard selection

Sparse BM25 formula (Publicus)

• Tokenizer: cvefixes-ascii-code-nfkc-casefold/v1 (Publicus HF sparse formula)
• Ranking params: k1=1.2, b=0.75, title_weight=5.0, body_weight=1.0
• Terms are sorted lexicographically; postings shards expose first_term/last_term
• IDF: FTS5 log((N - df + 0.5) / (df + 0.5)) with floor 1e-6
• Shard bounds: ≤ 4096 rows / ≤ 4096 terms per postings shard

Counts

• corpus documents: 2,174
• BM25 terms: 30,514
• BM25 postings: 854,046
• vectors: 2,174
• graph nodes: 37,764 (incl. BM25 terms: 30,514)
• graph edges: 883,628 (contains_term: 854,046)

Efficient retrieval

1. Load meta indexes under indexes/ (keyword shards, vector chunks, adjacency).
2. Probe BM25 by term range → download only matching data/bm25/postings/part-*.parquet.
3. Probe vectors by centroid → download only relevant vector shards.
4. Expand via graph adjacency pages; BM25 term nodes share the lexical vocabulary.

BM25 layout fragment:

{
  "average_document_length": 1771.7778288868444,
  "b": 0.75,
  "body_weight": 1.0,
  "k1": 1.2,
  "max_query_terms": 64,
  "posting_rows_per_record": 4096,
  "terms_per_shard": 4096,
  "title_weight": 5.0,
  "tokenizer": "cvefixes-ascii-code-nfkc-casefold/v1"
}

Primary identity

• Primary key: content ID (entry_cid / CIDv1 bafk…)
• Secondary pointer: dense document_index (0..N-1) used only inside sparse postings and shard ranges
• Graph endpoints use node_cid / edge_cid / subject_cid / object_cid

Coverage

Family	Documents	Source pin
CFR Title 37	1,246	cfr-title37-annual-2024 (through 2024-07-01)
MPEP	746	mpep-mpep-9-r07.2022 (through 2022-07-01)
35 U.S.C.	175	uscode-title35-2024
USPTO guidance	7	uspto-guidance-pdfs-2024-07-17
Total	2,174

Query client (local / e2e)

From the ipfs_datasets_py repository:

# Rule retrieval evaluation suite (BM25 + hybrid, citation-aware ranking)
python3 scripts/ops/legal_data/query_patent_legal_ir_hf.py \
  --release-dir /path/to/local/checkout-or-download \
  evaluate --strict

# Interactive BM25 (citation + topic→section boost ON by default)
python3 scripts/ops/legal_data/query_patent_legal_ir_hf.py \
  --release-dir /path/to/local/checkout-or-download \
  bm25 "duty of candor and good faith" --top-k 10

# Explicit section ranking
python3 scripts/ops/legal_data/query_patent_legal_ir_hf.py \
  --release-dir /path/to/local/checkout-or-download \
  bm25 "information disclosure statement 37 CFR 1.97"

Citation-aware re-ranking:

• explicit cues: 37 CFR 1.97, 35 USC 103, MPEP § 2001, bare 1.56
• topic lexicon: “duty of candor” → 1.56 / MPEP 2001; “IDS timing” → 1.97/1.98/609
• family-safe matching (avoids false friends like CFR 2.103 for USC 103)

Verified e2e (local): 9/9 rule cases PASS; pytest integration suite 13/13 PASS.

Snapshot roots

• corpus_root_cid: bafkreiak2bzrnblycry6t34kyusupep3nusxpptwlaffaz7glur3cpeiem
• release_root_cid: bafkreidjgszp2welhobw42nhsvxebcjt7uugqsbufgzph46ghvj2r6vose
• layout: publicus-ir-graphrag/v1
• schema: patent.public_legal_ir_hf_release/v1