Nepali Law Search Engine

Overview

This project is a Retrieval-Augmented Generation (RAG) system built to search and understand Nepali legal documents.
It combines traditional keyword search (BM25), semantic vector search (FAISS), and cross-encoder reranking to provide the most relevant results, and can generate answers using Gemini API.

Features

• Hybrid Search: Combines BM25 (Elasticsearch) and FAISS semantic similarity.
• Cross-Encoder Reranker: Reorders retrieved documents using a transformer-based reranker for higher accuracy.
• Multilingual Support: Handles both English and Nepali texts with multilingual models.
• RAG Integration: Generates answers using the retrieved context and Gemini LLM.
• Streamlit UI: Has a streamlit UI for the user

---

Project Structure

│
├── dataset/
|   ├── edited/
│   ├── indexes/
|   |   ├── faiss/
│   ├── processed/
│   ├── embeddings/
│   └── raw_pdfs/
├── notebooks/ 
├── scripts/
├── requirements.txt 
├── .gitignore 
└── README.md

How the Search Works

1. User Query: The user inputs a question or keyword (either in English or Nepali).
2. Document Retrieval: Both BM25 and FAISS independently retrieve top chunks from all law documents.
3. Reranking: The cross-encoder model assigns relevance scores and ranks them accordingly.
4. Context Formation: The top results are combined to form a structured context.
5. Answer Generation: The RAG model (using Gemini API) generates a natural-language answer based on the retrieved context.
6. Display: The Streamlit interface shows both the generated answer and the supporting references (chapter, section, law name).

Multilingual Handling

• The system supports both Nepali and English queries and documents.
• A multilingual sentence transformer is used for embeddings to ensure cross-language semantic understanding.
• Each document in the dataset includes:
  • doc_id (unique per law or language version)
  • chunk_id
  • chapter
  • section
  • text
• All documents (both languages) are stored under a single Elasticsearch index and included in the same FAISS index.

Pre-requisite

1. Elasticsearch 
2. Python(3.10+)

Installation

1.Clone the Repository

git clone https://github.com/fuseai-fellowship/Law-search-Engine-Nepali.git
cd Law-search-Engine-Nepali

2. Create and Activate Virtual Environment

python -m venv env
source env/bin/activate

3. Install Dependencies

pip install -r requirements.txt

4. Setup environment variable

ELASTIC_PASSWORD=your_elasticsearch_password 
GEMINI_API_KEY=google_studio_api_key

5. Start docker service

To setup elasticsearch indexing please visit our notebooks section

docker start elasticsearch_container

6. Run streamlit application

cd scripts/
streamlit run app.py

The app will be available at: 👉 http://localhost:8501

7. Demonstration

Future Enhancements

• Add context summarization for long documents
• Include Nepali Question Answering (QA) model fine-tuned on legal datasets
• Integrate document upload and indexing directly from UI
• Enhance search visualization and ranking transparency