legal-embedding-1.0

A sentence-transformers bi-encoder, fine-tuned from all-MiniLM-L6-v2 on contract clauses, that maps a clause to a 384-dimensional vector for semantic similar-clause search and retrieval-augmented contract Q&A. It is Model 4 of five in ContractGuard AI, an open-source clause-level contract analysis platform.

Not legal advice. This model retrieves clauses by semantic similarity. It does not itself determine legal effect — retrieved evidence is intended for a review workflow with a human in the loop.

Model summary

Base model	sentence-transformers/all-MiniLM-L6-v2
Output dimension	384
Task	Sentence/clause embedding for cosine-similarity retrieval
Training data	CUAD, 11,144 examples from 507 documents
Version	1.0 (active)
License	Other — see License

Intended use

Embed a contract clause and use cosine similarity against other embedded clauses for:

• Similar-clause search — surfacing comparable clauses from a user's own prior contracts.
• Retrieval-augmented contract Q&A — retrieving the clauses most relevant to a natural-language question before an answer is generated or assembled from evidence.

In scope: English-language commercial contract clauses, already segmented from the source document.

Out of scope: cross-lingual retrieval, whole-document embedding, and any comparison against vectors from a different model or version — embeddings are only meaningful compared against other embeddings from this exact model and version (see Model-version scoping).

How to use

from sentence_transformers import SentenceTransformer
from sentence_transformers.util import cos_sim

model = SentenceTransformer("AnuragBodkhe/legal-embedding-1.0")

clauses = [
    "Either party may terminate this Agreement upon thirty (30) days' written notice.",
    "This Agreement may be terminated by either party with 30 days prior notice.",
    "The Confidential Information shall not be disclosed to any third party.",
]

embeddings = model.encode(clauses, normalize_embeddings=True)
print(cos_sim(embeddings[0], embeddings[1]))  # near-duplicate termination clauses
print(cos_sim(embeddings[0], embeddings[2]))  # unrelated confidentiality clause

The embedding width (384) is a hard contract with the rest of the platform: the embeddings table, the service's embedding-dim setting, and this model's hidden size are all enforced to agree — a model emitting a different width is refused at load rather than silently corrupting retrieval.

Training data

CUAD contract clauses, split by document to prevent leakage:

Split	Examples	Documents
Train	7,705	353
Validation	1,711	70
Test	1,728	84

Training pair construction

Positive pairs are same clause type, from different documents. Pairing two clauses from the same contract was deliberately avoided — it teaches the encoder to recognise a drafting house style rather than a legal concept, which is not what similar-clause retrieval needs.

Evaluation mirrors that constraint: a query's own source document is excluded from its candidate pool entirely, so the model cannot be rewarded for retrieving a sibling clause it has effectively already seen.

Training procedure

Hardware	1× RTX 3050 6 GB (laptop GPU), fp16, CUDA 12.6
Wall clock	5 minutes

Evaluation results

Measured on the held-out test split (1,728 examples from 84 unseen documents):

Metric	Before fine-tuning	After fine-tuning
MRR	0.7412	0.8412
recall@5	0.0330	0.0465
precision@5	—	0.7682
nDCG@10	—	0.7631

What this changes in practice

The recall numbers understate the practical effect. Before fine-tuning, contract Q&A answered nothing for questions with no lexical overlap between the question and the clause text; after, it retrieves the right clause directly:

Question	Retrieved clause	Before fine-tuning
"How much notice must I give to terminate?"	7. TERMINATION	no evidence found
"Is there a non-compete?"	8. RESTRICTIVE COVENANT	no lexical overlap
"Who owns the copyright in my work?"	5. INTELLECTUAL PROPERTY	no lexical overlap

Limitations and bias

• Training population. CUAD skews toward US commercial contracts. Retrieval quality on other jurisdictions or contract families has not been separately evaluated.
• Not a substitute for exact matching. For near-duplicate or boilerplate-variant clauses, a lexical baseline (signed hashing over word unigrams, bigrams and character 4-grams) can perform comparably or better; this model's advantage is on paraphrase, where lexical overlap is low or absent.
• A low-similarity result means "the contract does not address this," not a best guess. The serving platform withholds results below a cosine similarity of 0.25 and says so explicitly, rather than returning the nearest vector regardless of how far it is.
• Model-version scoping. Every similarity query in the serving platform filters on model name and version. Comparing vectors from two different model versions returns confident, silently wrong results — the platform's embedding store enforces this filter for exactly that reason, and any downstream use of this model outside that platform should do the same.

Index recommendation

For production-scale retrieval, an HNSW index over cosine distance is recommended (used as-is in ContractGuard's Postgres/pgvector deployment):

CREATE INDEX embeddings_vector_idx ON embeddings
    USING hnsw (embedding vector_cosine_ops) WITH (m = 16, ef_construction = 64);

Order by the distance operator directly rather than a derived similarity score — only the operator form can use the index.

License

The base model, sentence-transformers/all-MiniLM-L6-v2, is licensed Apache-2.0. Training data license: CUAD (CC BY 4.0). Review the upstream license terms before commercial use in combination with this model's own license terms.

Part of ContractGuard AI

This model is one of five in the ContractGuard AI pipeline (clause classification, risk classification, legal NER, retrieval embeddings, contract NLI). See the main repository and docs/RETRIEVAL.md for the full retrieval design, including the dimension contract, model-version scoping, and user scoping enforced around this model.