Bulk Data

The U.S. caselaw corpus, wholesale

10 million state and federal opinions — the standard case set the legal industry works from — with the citation graph, U.S. Code, federal rules, full-text index, and semantic vectors that make it queryable. Licensed for direct database access, replication into your own environment, or API integration.

What the corpus contains

Every layer below is a queryable store in its own right, not a derived export. The AI analysis layer carries its own vector set, embedded separately from the opinion text, so meaning-based retrieval runs against both.

Opinions

10 million rows

One row per opinion — majority, concurrence, and dissent each addressable — with normalized court, jurisdiction tier, filing date, docket, publication status, and full body text.

Citation graph

78 million edges

Resolved opinion-to-opinion edges, traversable in both directions: what a case cites, and everything that has cited it since.

In-text citations

182 million rows

Every citation as it appears in the prose, with character offsets, pin cite, and the short-form/supra/id resolution chain — extracted by GoodLaw (built on eyecite, then extended in-house) across 10 million completed scans, with an LLM pass for what the deterministic parser misses.

Reporter citations

18 million rows

Parallel citations with a normalized reporter key per row, so official, regional, and vendor cites all resolve to the same decision.

Full-text index

10 million docs / 133 GB

Apache Lucene inverted index (OpenSearch) over the whole corpus, with per-field boosting, positional offsets for phrase and proximity queries, and BM25 relevance scoring.

Semantic vectors

470,000 vectors

1024-dimension Qwen3-Embedding-8B embeddings, stored as pgvector halfvec and indexed with binary-quantized HNSW. Two sets are built separately and searched together: AI analysis embeddings and opinion-body embeddings.

AI analysis

56,000 opinions

Model-authored issue, holding, posture, and quotation extraction — with its own dedicated vector set, embedded and indexed separately from the opinion body, for higher precision on hard-to-find caselaw. Authored by LawDiver; not derived from Westlaw, Lexis, or Shepard’s.

United States Code

25 years

The app is wired directly into twenty-five years of the U.S. Code, fully accessible and cross-indexed in both directions: every case that cites a code section, and every code section that is cited by a case.

Federal rules

Full federal sets

The Federal Rules of Civil Procedure, Criminal Procedure, Evidence, and Appellate Procedure, plus related federal rule sets, stored as addressable rules and subdivisions and linked to the opinions that cite them.

Code of Federal Regulations

25 years (wiring now)

Twenty-five years of the C.F.R., cross-indexed as cited-to and cited-by with caselaw. This layer is being wired now and is not yet visible in the public research UI; it is part of the same corpus architecture as the U.S. Code.

How it is built

Four purpose-built stores rather than one. Retrieval by citation, by name, by keyword, and by meaning are different physical problems, and each gets the index that suits it.

Primary corpus — PostgreSQL 18

The 303 GB store of record: opinions, the citation graph, in-text citations, reporter citations, and the AI analysis tables. Relational and queryable, not a document dump.

Citation index — dedicated Postgres store

15.5 million rows keyed on a normalized reporter key, isolated in its own database so a lookup by citation is a single indexed hit rather than a scan of the main corpus.

Name index — dedicated Postgres store

10 million rows with trigram (GIN/pg_trgm) indexes on case name and both party names, for fuzzy retrieval by name that tolerates abbreviation and misspelling.

Full-text search — OpenSearch / Apache Lucene

Stemmed, ASCII-folded analysis with positional offsets, supporting boolean, phrase, and proximity queries with BM25 ranking and field-level boosts across body, name, and syllabus.

Semantic search — pgvector + OpenSearch k-NN

470,000 Qwen3-Embedding-8B vectors at 1024 dimensions, stored as halfvec and indexed by binary-quantized HNSW (Hamming first pass, full-precision rerank) for cheap recall at scale.

Two vector sets, not one

The AI analysis layer carries its own embeddings, built separately from the opinion-body embeddings and searched alongside them. A query phrased as a legal question matches text already written in the language of issues and holdings, which is what surfaces relevant caselaw that keyword and body-only vector search miss.

Citation extraction — GoodLaw

Every opinion is scanned with LawDiver’s proprietary GoodLaw system, built on the open-source eyecite parser and substantially extended in-house. GoodLaw is deterministic: the same opinion yields the same cites. An LLM pass then searches for anything GoodLaw missed, so reliability and completeness are stacked rather than traded off.

How you get it

Three access models, all reading the same continuously updated corpus.

Direct database access

Credentialed, read-only connections to the Postgres stores and the OpenSearch cluster on our infrastructure. You issue your own SQL and your own queries — no export step, no staleness.

Replica on your infrastructure

The corpus stood up inside your environment as a mirror, seeded from a bulk snapshot and kept current by incremental sync. Appropriate where data cannot leave your network.

REST API

The same corpus over HTTP, for retrieval, search, and citation lookup. No request quota — throughput is bounded by server and network capacity, negotiated per agreement.

Daily incremental updates

New decisions are ingested continuously, normalized, citation-scanned, and propagated across every store and index — so all delivery methods converge on the same current state.

Statutes, rules, and regulations

Caselaw is only one layer. The same graph ties opinions to the code, the federal rules, and — as wiring completes — the C.F.R.

United States Code — 25 years

LawDiver is wired directly into twenty-five years of the U.S. Code. The text is fully accessible. Every citing relationship is stored both ways: all cases that cite a given section, and all sections cited by a given case.

Federal rules

The federal rules of civil procedure, criminal procedure, evidence, and appellate procedure — and related federal rule sets — are stored as first-class authority, subdivided the way lawyers cite them, and linked to the opinions that invoke them.

Code of Federal Regulations — 25 years

The same cited-to / cited-by design is being applied to twenty-five years of the C.F.R. You cannot see this layer in the public app yet; it is being wired now into the same stores and indexes as the U.S. Code.

Where the data comes from

The corpus is built from public sources only. The backbone is the quarterly CourtListener bulk downloads, which are available without restriction. Around that, LawDiver harvesters run continuously against more than 200 individual court websites, public RSS feeds of newly issued opinions, and the massive GovInfo site (govinfo.gov), which publishes federal statutes, regulations, and caselaw.

Those streams are normalized into one schema, citation-scanned, and folded into the graph, the full-text index, and the vector stores. A longer account of harvest, indexing, embeddings, and GoodLaw — without the licensing terms — is on Sources and Methods.

No licenses that constrain use

LawDiver represents and warrants that the entire LawDiver database can be used without violating any licenses, copyrights, or terms of service. The corpus was assembled without accepting any license, terms of service, or other restriction that would constrain how LawDiver uses the data, or how LawDiver customers are able to use the data.

All caselaw analysis in LawDiver was created by LawDiver. Nothing was taken from other research services. Westlaw headnotes were not used. Lexis headnotes were not used. Shepard’s was not used.

Citation extraction and treatment use LawDiver’s proprietary GoodLaw system. GoodLaw started from the open-source eyecite application and was greatly enhanced by our in-house team to recognize the full range of legal citations and to deliver far more complete results. A deterministic GoodLaw pass is followed by an LLM pass so that nothing the parser misses is left on the table.

Coverage and provenance

Coverage is U.S. Supreme Court, federal courts of appeals, federal district courts, and the appellate and supreme courts of all fifty states. Opinions are sourced from the public record and normalized into a single schema: one row per opinion, so a dissent is addressable separately from the majority it accompanies, with court, jurisdiction tier, circuit, filing date, docket number, and publication status resolved to controlled values rather than left as source strings.

Model-authored fields are stored in their own tables, physically separate from court-authored text. Nothing a model wrote is ever mixed into an opinion's body, syllabus, or headnotes — so an evidentiary distinction that matters to a court is structural in the schema rather than a naming convention.