Legal embeddings
Turns a legal query into a dense vector that carries meaning, citation, and doctrine, so the right authority surfaces rather than the nearest keyword match.
Foundational AI · For the law
Foundational AI,
Foundational AI,
trained on the law.
BRELYNE develops foundational models purpose-trained for legal text and the tasks that turn on every word: retrieval that surfaces the controlling authority, enrichment that renders documents as structured graphs, and the open infrastructure beneath both.
The thesis
General models approximate the law.
Ours are trained on it.
BRELYNE is a research-led developer of foundational AI models built for one domain and held to one standard. We do not adapt general-purpose models to legal work. We train models on legal text and the tasks that decide a matter, then hold them to the accuracy that practice actually requires.
Where other systems treat law as one more genre of writing, we treat it as the discipline it is: citation, doctrine, structure, and consequence. Every model we ship began as a problem our team lived in legal practice and legal technology, from pulling a single date out of a contract to finding the one decision that controls somewhere across thousands of cases.
The models
A small, deliberate family
Four models, each purpose-built for a stage of legal work: retrieval that finds the right authority, enrichment that renders a document as structure, and the open layer that makes both more accurate.
Legal reranking
Orders a field of candidates by true legal relevance, weighing how directly a passage controls, distinguishes, or merely mentions the question at hand.
Graphitization
Transforms documents of any length into structured knowledge graphs of parties, clauses, citations, and holdings, at sub-second latency across a corpus.
Semantic chunking
The open-source layer beneath the rest. Splits documents where meaning breaks rather than where a character count happens to fall. More than two million downloads a month.
By the numbers
What the work measures
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.
Deployment
Run it on your terms.
The same models and the same API, reached three ways. From fully managed to fully air-gapped, the model is identical and only the perimeter changes.
Platform
Hosted APIs you can call today, built for production legal workloads with no permanent data retention.
Marketplace
Procure through a major cloud marketplace, on the billing and security relationship you already trust.
Self-hosted
Air-gapped containers inside your own cloud tenancy or on your own hardware. Nothing leaves your environment.
What the models do
Legal retrieval, embeddings, reranking, knowledge graphs, graphitization, semantic chunking, citation search, contract analysis, date extraction, case law, statutes, decisions, document enrichment.
The work ranges from the deceptively simple to the genuinely hard. Pull every effective date from a stack of contracts. Find the one case that controls somewhere in thousands. Map a filing into its parties, claims, and authorities, then ask it questions. We build for the people who do this work: legal-technology companies and the law firms they serve.
For developers
Build on the same models
The models behind our platform are available to build with directly. Call the embedding and reranking APIs, graphitize a document, or wire the open chunker into your pipeline. The interface stays the same whether you run it hosted, through a cloud marketplace, or inside your own air-gapped containers.
# Embed and rerank a legal query POST https://api.brelynetech.net/v1/retrieve { "query": "duty to warn of latent defect", "corpus": "us-case-law", "top_k": 3, "rerank": true } # 200 OK · 38 ms · retention: none { "results": [ { "rank": 1, "score": 0.94, "cite": "22 U.S. 1" }, { "rank": 2, "score": 0.89, "cite": "17 U.S. 316" } ], "status": "ranked" }
Illustrative request and response. Citations shown are public landmark authorities.
Start with a conversation.
Tell us the matter you are trying to win or the product you are trying to ship. We will route you to the right models, the right deployment, and the credits to begin.