Technology

A family of models, built for legal text.

Four models cover the arc of legal work: chunking that respects meaning, embeddings and reranking that find the controlling authority, and graphitization that renders a document as structure. Each is trained on legal text from the start, and each is measured against the accuracy the work demands.

Model 01 · Retrieval

Legal embeddings

An embedding model trained to read law the way practitioners do. It turns a complex legal query into a dense vector that captures meaning, citation, and doctrine, so the right authority surfaces and not merely the right keywords. Where general-purpose embeddings flatten a statute and a sales contract into the same fuzzy notion of a document, ours hold the distinctions that decide a matter.

Paired with the reranker, it forms what we present as the most accurate legal retrieval pair available, ranking among the top performers on independent legal retrieval and embedding benchmarks. It is the difference between a search that returns plausibly related text and one that returns the case you needed.

Vector searchCitation-awareBenchmark-leading

Query, mapped to vector space, nearest authorities

Model 02 · Retrieval

Legal reranking

A reranking model that takes a field of candidate results and orders them by true legal relevance, weighing how directly a passage controls, distinguishes, or merely mentions the question at hand. Retrieval gets you close. Reranking decides which of the close calls actually belongs at the top.

Together, embedding and reranking turn complex legal queries into highly relevant results, supporting tasks such as searching large volumes of cases for a single critical citation. They are designed as a system: recall you can trust, ordered by precision you can defend.

Relevance orderingCross-encoderPrecision at rank 1

Candidates, reranked, controlling authority first

Model 03 · Enrichment

Graphitization

A hierarchical document-enrichment model that transforms unstructured documents of any length into structured knowledge graphs. Laws, decisions, and contracts become connected entities: clauses, parties, dates, courts, holdings, and citations. It reads a document the way a careful associate would, then writes down the structure they would have held in their head.

It does this at sub-second latency, so enrichment is something you run across a corpus rather than schedule overnight. The result is law you can traverse: query a clause, follow a citation, pull every party, as a graph rather than a wall of text.

Knowledge graphsAny lengthSub-second latency

Unstructured document, rendered as a knowledge graph

Model 04 · Infrastructure

Semantic chunking

Beneath the models sits a semantic text-chunking algorithm we created and open-sourced. It splits documents where meaning actually breaks, at the seam between an argument and its exception or a clause and its proviso, rather than where a character count happens to fall. Good retrieval starts with good chunks, and ours keep an idea intact.

It has become widely used infrastructure, adopted by several of the largest technology and software organizations, with more than two million monthly open-source downloads. It is the quiet layer that makes everything above it more accurate.

Open sourceSemantic boundaries2M+ downloads a month

Split where meaning breaks, not where bytes do

How it composes

One pipeline, four stages

The models are built to work as a system. A document enters as raw text and leaves as something you can search with confidence and traverse as structure.

Stage 01

Chunk. Semantic chunking splits the source at natural boundaries, keeping each argument, clause, and proviso whole so nothing meaningful is cut in half.

Stage 02

Embed. Legal embeddings map every chunk and query into a shared vector space that carries citation and doctrine, not just surface wording.

Stage 03

Rerank. The reranker reorders candidates by how directly each one controls the question, putting the decisive authority first.

Stage 04

Graphitize. Graphitization renders the retrieved documents as connected entities, so a result is a graph you can query rather than a page you have to read.

Purpose-trained vs. general-purpose

The difference the domain makes

General-purpose models can read legal text. They were not built to be right about it. The table below is how we think about the gap, and what we designed the model family to close.

A comparison of general-purpose models and BRELYNE across key dimensions of legal retrieval and enrichment.
Dimension	General-purpose models	BRELYNE
Training focus	Broad web and general text	Legal text, purpose-trained for the domain
Retrieval unit	Fixed character or token windows	Semantic chunks that keep an idea intact
What surfaces	Text that looks related	The authority that actually controls
Document structure	Flat passages of text	Knowledge graphs of parties, clauses, and citations
Latency at scale	Batch or overnight enrichment	Sub-second, run across a whole corpus
Deployment	Vendor cloud only	Hosted, marketplace, or air-gapped self-hosted

By the numbers

Measured against the work

2M+

Monthly open-source downloads

Semantic chunking algorithm

~26%

More accurate

vs. general-purpose equivalents

~30%

Faster inference

vs. the next-best alternative

Figures reported by BRELYNE. Our legal retrieval models rank among the top performers on independent legal retrieval and embedding benchmarks.

In development

A world of law, structured.

A forthcoming offering gives direct access to a proprietary repository of laws, decisions, contracts, and other legal documents from around the world, already graphitized and ready for the models to reason over. The corpus and the models, built for each other.

Request early access