| --- |
| language: |
| - en |
| license: apache-2.0 |
| library_name: sentence-transformers |
| pipeline_tag: sentence-similarity |
| tags: |
| - legal |
| - legal-retrieval |
| - bns |
| - indian-law |
| - semantic-search |
| - information-retrieval |
| - sentence-transformers |
| --- |
| |
| # CourtInsight BNS Encoder |
|
|
| CourtInsight BNS Encoder is a BNS-specific semantic embedding model developed for the CourtInsight legal-information retrieval system. |
|
|
| The model is designed to retrieve relevant provisions of the Bharatiya Nyaya Sanhita, 2023 (BNS) from natural-language descriptions of legal situations. |
|
|
| The retrieval approach is semantic and does not use keyword matching, BM25 retrieval, manually defined keyword lists, or hard-coded section mappings. |
|
|
| ## Intended Use |
|
|
| - Semantic search over BNS 2023 |
| - BNS provision retrieval |
| - Legal-information retrieval |
| - Retrieval-Augmented Generation (RAG) |
| - Legal education and research applications |
| - Natural-language search interfaces |
|
|
| The model is not a legal-advice system and should not be used to determine legal liability, make judicial decisions, or replace qualified legal professionals. |
|
|
| ## Base Model |
|
|
| `nomic-ai/nomic-embed-text-v1.5` |
|
|
| The model is compatible with Sentence Transformers. |
|
|
| For queries, use the prefix: |
|
|
| `search_query:` |
|
|
| For documents, use: |
|
|
| `search_document:` |
|
|
| ## Training |
|
|
| The training corpus was constructed from a structured dataset covering 358 BNS 2023 sections. |
|
|
| The dataset contains statutory information including section number, section title, operative text, explanation, illustration, legal elements, and element summaries. |
|
|
| Semantic hard-negative mining was performed using the base Nomic embedding model. |
|
|
| The V3 training dataset contained 10,545 query-positive-negative triplets. |
|
|
| Training used TripletLoss with cosine distance. |
|
|
| ## Training Configuration |
|
|
| | Parameter | Value | |
| |---|---| |
| | Base model | nomic-ai/nomic-embed-text-v1.5 | |
| | Loss | TripletLoss | |
| | Distance | Cosine distance | |
| | Triplet margin | 0.5 | |
| | Epochs | 5 | |
| | Effective batch size | 16 | |
| | Learning rate | 2e-5 | |
| | Maximum sequence length | 512 | |
| | Precision | FP16 | |
|
|
| ## Evaluation |
|
|
| ### Controlled Natural Benchmark |
|
|
| 15 manually constructed legal scenarios: |
|
|
| - Top-1: 86.67% |
| - Top-3: 93.33% |
| - Top-5: 100% |
| - MRR: 0.9167 |
|
|
| These results use the complete CourtInsight retrieval pipeline. |
|
|
| ### 100-Query Coverage Evaluation |
|
|
| - Top-1: 99.00% |
| - Top-3: 100.00% |
| - Top-5: 100.00% |
| - MRR: 0.9950 |
|
|
| This benchmark is statute-derived and should not be interpreted as real-world user-query accuracy. |
|
|
| ### 1,074-Query Held-Out Evaluation |
|
|
| - Top-1: 98.23% |
| - Top-3: 99.72% |
| - Top-5: 99.72% |
| - MRR: 0.9899 |
|
|
| This benchmark is synthetic/title-derived and should not be interpreted as real-world legal accuracy. |
|
|
| ## CourtInsight Architecture |
|
|
| User Query |
|
|
| → CourtInsight BNS Encoder |
|
|
| → Top-10 Semantic Candidates |
|
|
| → Cross-Encoder Reranking |
|
|
| → Nomic + Cross-Encoder Score Fusion |
|
|
| → Top-3 BNS Provisions |
|
|
| The retrieval pipeline does not use keyword matching, BM25, or hard-coded section mappings. |
|
|
| ## Example |
|
|
| ```python |
| from sentence_transformers import SentenceTransformer |
| |
| model = SentenceTransformer( |
| "Ayushi054/CourtInsight-BNS-Encoder", |
| trust_remote_code=True |
| ) |
| |
| query = "A person knowingly uses a forged document as genuine." |
| |
| embedding = model.encode( |
| "search_query: " + query, |
| normalize_embeddings=True |
| ) |
| ``` |
|
|
| ## Limitations |
|
|
| - Specialized for the Bharatiya Nyaya Sanhita, 2023. |
| - English-language retrieval. |
| - Not a case-law retrieval model. |
| - Ambiguous or incomplete queries may retrieve closely related provisions. |
| - Similar statutory provisions can be difficult to distinguish. |
| - Evaluation results do not establish real-world legal accuracy. |
| - Retrieved provisions should be independently verified against the applicable legislation. |
|
|
| ## License |
|
|
| Apache-2.0 |
|
|
| ## Disclaimer |
|
|
| CourtInsight is an experimental/research legal-information retrieval system. It is intended to help users locate potentially relevant BNS provisions. It does not provide legal advice and should not be treated as a substitute for a qualified legal professional or authoritative legal source. |
