法律检索库

LLM Law Library

Open legal research infrastructure for humans and AI. It retrieves real US court opinions, statutes, statutory-definition pointers, and court rules without a subscription. Retrieval does not replace legal judgment: a real citation can still be irrelevant, mischaracterized, stale, or attached to noncontrolling text.

Free. Open. Primary-source grounded.

Only have a chat window and web access? Open For Chat-Window LLMs. It contains complete GET URLs and usable research recipes; no tool installation or skill loader is required.

What's In The Box

Data Sources

• Caselaw: CourtListener / Free Law Project. CC0 public domain.
• Statutes: Scraped from official state legislative websites + OLRC USLM XML for federal. Public law.
• Embeddings: all-MiniLM-L6-v2 (384-dim), pgvector IVFFlat.
• BM25: ParadeDB pg_search (Tantivy/Rust).

Access

MCP Server (Claude Code / Claude Desktop)

The llm-law-library MCP server exposes 12 tools:

URL-only GET API

Every research read is available as an ordinary GET URL for capable models that can browse the web but cannot install MCP or send POST bodies. See For Chat-Window LLMs for copyable URLs and complete research plans.

JSON POST API

Base URL: https://implausible.enterprises/api/caselaw/

Web UI

https://implausible.enterprises/projects/law-library/

Key Concepts

Shepardizer

Named after Shepard's Citations (1873). Tracks how later courts treated earlier decisions. Built from the CourtListener citation graph: 3.7M classified treatments across 16 granular types.

Severity hierarchy: overruled > abrogated > reversed > vacated > superseded > questioned > distinguished > limited > criticized > modified

Treatment fields are automated leads derived from citation-graph and opinion-text analysis. They can be noisy or misattributed and must not be reported as a good-law conclusion without inspecting the later opinions and checking an authoritative citator where the stakes require it.

Red Flags (Used-But-Undefined)

The most dangerous gap in legal research is a definition you don't know exists. A lawyer briefs the ordinary meaning of a term, but the statute defines it to include or exclude key categories.

Red flags detect this: a term is heavily used in a jurisdiction's statutes but has no statutory definition found there. Four possible causes:

1. Cross-reference gap — definition exists in a general definitions chapter we haven't linked
2. Cross-reference we missed — "fiduciary has the meaning given in 15 U.S.C. § 80a-2"
3. Common-law reliance — legislature intentionally uses common-law meaning (legitimate, worth flagging)
4. Recall gap — we missed something in extraction

All four are worth surfacing. Cases 1-2 are actionable; case 3 is informative; case 4 improves the system.

Definition Pointers

We store where terms are defined, not the definitions themselves. "Person" might have 40 different statutory definitions across 30 jurisdictions, each with different scope ("as used in this chapter," "for purposes of this title," etc.). We point you to the right section; you read it in context.

Signal types (trust hierarchy, highest first):

• xml_structured — from authoritative XML source (federal USLM)
• caselaw_cite — found in court opinions citing the statute as defining this term
• bm25_pattern — found via BM25 phrase search (linking phrases: "means," "shall include," etc.)
• flash_classification — LLM-classified definition section
• red_flag — term used but not defined in this jurisdiction

Search Modes (Caselaw)

Determinism matters for auditability. BM25 makes the same lexical query reproducible; that is useful evidence of what was searched, not a guarantee that the search was complete or legally adequate.

Jurisdiction Codes

• Federal: federal
• States: state:XX where XX is the two-letter code (e.g., state:CA, state:NY, state:TX)
• DC: state:DC

All 50 states + DC + federal are loaded and searchable.

Architecture

Client (MCP/API/Web)
  -> fauna-box API (port 8030, LRU cache, CORS)
  -> HTTP -> Mac server (port 8031, embedding model in memory)
  -> PostgreSQL 17 + pgvector + ParadeDB pg_search

Two-box architecture: fauna-box runs the API/proxy/MCP layer; Mac hosts the PostgreSQL corpus with pgvector and ParadeDB.

Database Tables

Caselaw

• rag_passages — 76.9M passages with 384-dim embeddings
• opinion_search — 10.9M opinions, denormalized BM25 index
• shepardizer_treatment_v2 — 29.5M treatment rows

Statutes

• statute_search — 2.34M sections with BM25 index, all 52 jurisdictions

Definitions

• definition_pointers — 882K rows, where terms are defined
• term_registry — 215K unique legal terms
• extracted_definitions — 715K extracted definition records
• definition_embeddings — 135K definition embeddings for alias discovery
• statute_embeddings — 1.79M statute section embeddings
• term_aliases — 160 LLM-confirmed cross-state term aliases

Scripts

Shepardizer v2

• scripts/shepardizer-v2-pass1.py — Citation graph + regex treatment classification
• scripts/shepardizer-v2-pass2.py — Haiku verification of treatments
• scripts/shepardizer-v2-gemini-grinder.py — Gemini Batch verification (50% cheaper)
• scripts/shepardizer-driver.py — Autonomous verification driver

Definition Discovery

• scripts/statutes/bm25_definition_seed.py — BM25 phrase search across statute corpus
• scripts/statutes/batch_classify_all.py — Gemini Batch API classification
• scripts/statutes/batch_extract_all.py — Gemini Batch API extraction
• scripts/statutes/batch_red_flags_all.py — Used-but-undefined scan (SQL only, no LLM)
• scripts/statutes/{xx}_full_cook/ — Per-state processing directories

BM25

• scripts/bm25/01_schema.sql — opinion_search table + extraction function
• scripts/bm25/03_index.sql — BM25 index with field boosts
• scripts/bm25/run_eval.py — Evaluation harness (30 queries)

License

Caselaw corpus: CC0 (CourtListener/Free Law Project). Statutory text: Public law, not copyrightable. Code: This project.