Lawreview research

Law Review Research

Use the bundled read-only CLI to search a compatible SQLite/FTS corpus. The skill does not include article text or a database; the user must supply material they are authorized to process.

Non-negotiable rules

• Treat OCR, title guesses, author guesses, journal hints, and year guesses as provisional.
• Expand context before relying on a search hit, then verify important language against an authoritative copy.
• Report corpus scope, filters, query mode, and known gaps. Do not describe silence in the corpus as absence in the literature.
• Quote only as much text as the user's rights and applicable law permit.
• Open databases read-only. Do not modify, rebuild, or redistribute a corpus unless the user explicitly authorizes it and has the necessary rights.
• Do not expose local paths, document identifiers, or corpus contents that the user has not authorized for disclosure.

Configure

Set the database path before running the tools:

export LAWCORPUS_DB=/path/to/lawreview_search.sqlite

Optionally configure a consistent snapshot used when the primary database is locked:

export LAWCORPUS_SNAPSHOT=/path/to/lawreview_search.latest.sqlite

Command-line --db and --snapshot-fallback values override the environment. Read references/schema.md when creating or diagnosing a compatible database.

Workflow

1. Run stats to understand the indexed corpus.
2. Use chunk search for normal research questions and passage retrieval.
3. Use context around promising chunks.
4. Use page search when the page location matters or chunking missed a phrase.
5. Use citing or citation mode for legal citation strings.
6. Fetch a document only when larger OCR context is necessary.
7. Verify important results against authoritative copies and report search limitations.

Run from this skill directory:

python scripts/lawcorpus.py stats
python scripts/lawcorpus.py search --mode chunk --limit 10 \
  '"reasonable person" objective negligence'
python scripts/lawcorpus.py context --chunk-id 9795 --before 2 --after 2
python scripts/lawcorpus.py citing '15 U.S.C. § 1' --limit 25

Useful options:

search --mode chunk|page|citation
       --limit N
       --journal TEXT
       --after YEAR
       --before YEAR
       --allow-duplicate-documents

document --document-key KEY | --source-id ID | --member-like TEXT
         --max-chars N

context  --chunk-id ID
         --before N --after N

Chunk and page search return at most one hit per document by default. Allow duplicates only when exhaustive within-document retrieval matters more than breadth.

SQLite FTS5 syntax applies. Prefer direct legal terms and quoted phrases. If punctuation causes an FTS error, simplify the query; use citing for citation strings containing symbols such as §.

Evaluate retrieval

Use a JSONL gold set:

{"id":"privacy-1","query":"privacy tort appropriation likeness","mode":"chunk","expected_title_contains":["privacy"]}
{"id":"antitrust-1","query":"15 U.S.C. § 1","mode":"citation","expected_citation_contains":["15 U.S.C."]}

Run:

python scripts/evaluate_search.py queries.jsonl

Report recall@k and mean reciprocal rank before claiming a retrieval change improved results.

Optional MCP server

The bundled dependency-free JSON-RPC stdio server exposes the same read-only primitives:

python scripts/lawcorpus_mcp.py --db "$LAWCORPUS_DB"

Tools: corpus_stats, list_journals, search, get_document, get_context, and find_citing. Prefer the CLI for shell-based agents and MCP only when the runtime is configured to launch local stdio servers.

Deliver results

For each important hit, provide the provisional title/author/year when present, journal hint, passage or citation match, page/chunk location, and verification status. End with a compact coverage note identifying the query, mode, filters, corpus date or snapshot, and material limitations.