MyPocketLawyer ⚖️ — AI-Powered Legal Aid Assistant (Nepali Law)

A stateless Retrieval-Augmented Generation (RAG) assistant for Nepali law. It uses:

• Google Gemini for embeddings and generation
• ChromaDB for vector search
• FastAPI for the backend
• Vite/React and Shadcn/UI for the modern frontend user interface

This bot answers only from the ingested legal documents:

• Constitution of Nepal 2072
• The Criminal Offences Act 2074
• The Labour Act 2074
• The National Civil Act 2074
• The National Penal Act 2074
• Bank and Financial Institution Act 2073
• Banking Offence and Punishment Act 2064
• Electronic Commerce Act 2081
• International Financial Transactions Act 2054
• The Income Tax Act 2058

---

✨ Features

• Small-talk guard and domain classification (Nepali law only)
• Query rewriting to improve retrieval quality
• ChromaDB retrieval with metadata-rich source display
• Gemini 2.5 Pro/Flash for grounded answer generation
• Streamlit chat UI with collapsible sources and search query display
• Adaptive answer format:
  • Short Answer
  • What the Law Says (cites articles/clauses)
  • Practical Steps (only for action/procedure queries)
  • Disclaimer

---

🗂️ Repository Structure

.
├─ backend/
│  ├─ main.py                         # FastAPI stateless RAG pipeline
│
├─ chroma_db/                         # ChromaDB persistent directory (auto-generated)
│
├─ config/
│  ├─ __init__.py
│  └─ paths.py                        # Project path utilities (optional)
│
├─ data/
│  ├─ raw/                            # Raw legal documents (if building vectors)
│  ├─ processed/                       # Converted & cleaned JSON docs
│  └─ evaluation/                      # Test sets, metrics, prompts
│
├─ documentation/
│  ├─ LiteratureReview.pdf
│  └─ Proposal.pdf
│
├─ frontend/
│  ├─ index.html
│  ├─ package.json
│  ├─ vite.config.ts
│  ├─ tailwind.config.ts
│  ├─ postcss.config.js
│  ├─ public/
│  └─ src/
│     ├─ main.tsx                      # React entry point
│     ├─ App.tsx                       # Root wrapper component
│     ├─ App.css                       # Global component styling
│     ├─ index.css                     # Tailwind + base styles
│     │
│     ├─ pages/
│     │  ├─ Index.tsx                  # Landing page (hero + CTA)
│     │  └─ NotFound.tsx               # 404 handler
│     │
│     ├─ components/
│     │  ├─ Hero.tsx                   # Homepage hero UI
│     │  ├─ NavLink.tsx                # Navbar interactive link component
│     │  ├─ LegalAssistant.tsx         # Main chat screen
│     │  └─ ui/                        # shadcn component library
│     │
│     ├─ hooks/
│     │  ├─ use-mobile.tsx
│     │  └─ use-toast.ts
│     │
│     └─ lib/
│        └─ utils.ts                   # Shared helpers (stream UI, formatting)
│
├─ notebooks/
│  ├─ baseline_data_ingestion_pipeline.ipynb
│  ├─ baseline_retrieval_pipeline.ipynb
│  ├─ data_ingestion_multimodel.ipynb
│  ├─ final_data_ingestion_pipeline.ipynb
│  ├─ final_retreval_pipeline.ipynb
│  ├─ final_retriever_evaluation.ipynb
│  ├─ generator_evaluation.ipynb
│  ├─ multimodel_evaluation.ipynb
│  ├─ multimodel_split_evaluation.ipynb
│  └─ retriever_evaluation.ipynb
│
├─ .env                                # GEMINI_API_KEY
├─ .gitignore
├─ generator_evaluation_results.csv
├─ generator_evaluation_results.json
├─ requirements.txt
├─ run.py                              # Optional pipeline runner
└─ README.md

---

🧱 Tech Stack

• Backend: FastAPI, Pydantic, Uvicorn
• Vector Store: ChromaDB (persistent)
• LLM + Embeddings: Google Gemini (via google-genai)
  • Generation: gemini-2.5-pro (fallback: gemini-2.5-flash)
  • Embeddings: models/text-embedding-004
• Frontend: Frontend — Vite + React + TypeScript + Tailwind + Shadcn/UI

---

🚀 Quickstart

1) Prerequisites

• Python 3.10+
• A Google Gemini API key (from Google AI Studio)
• macOS, Linux, or Windows

2) Environment

Create a .env in the repo root:

GEMINI_API_KEY=your_gemini_api_key_here

3) Install dependencies

Create a virtual environment and install:

python -m venv .venv
# Windows: .venv\Scripts\activate
# macOS/Linux:
source .venv/bin/activate
pip install -r requirements.txt

4) Build or provide the vector store

• If ./backend/chroma_db/ already contains a collection named legal_docs, you’re set.
• Otherwise, build it using your ingestion notebooks:
  • Recommended: notebooks/final_data_ingestion_pipeline.ipynb
• Place raw text/markdown under data/raw/ (or what your notebook expects).
• Output should be a persistent Chroma collection at ./backend/../chroma_db (repo root ./chroma_db).

The backend looks for a Chroma collection named legal_docs at repo-root ./chroma_db.

5) Run the backend

From repo root:

uvicorn backend.main:app --host 0.0.0.0 --port 8000 --reload

You should see: “✅ MyPocketLawyer backend (completely stateless) is live.”

6) Run the frontend 🌐

In a separate terminal, follow these steps to launch the React application:

1. Navigate to the frontend directory

cd frontend

2. Install Node.js dependencies (only needed the first time)

npm install

3. Start the development server

npm run dev

Optional: run.py

If run.py orchestrates both backend and frontend on your machine, run:

python run.py

(If not implemented to launch both, use the separate commands above.)

---

🧪 Example Queries

• What are the three organs of the Government under the Constitution of Nepal?
• What is the overtime pay in Nepali like?
• What happens if someone is found planting explosives?

📊 Evaluation Figures

Retriever Evaluation (Full-Chunk Splitting)

Retriever Evaluation (Article-wise Splitting)

Generator Evaluation

Quality & Efficiency Scores

Response Time & Cost Distribution

Quality Metrics Comparison

Performance by Legal Sources