⚖️ Pakistan Law Chatbot — RAG System

A document-grounded AI chatbot that answers questions from uploaded Pakistani law PDFs using Retrieval-Augmented Generation (RAG), LangChain, Hugging Face embeddings, ChromaDB, and Ollama.

---

📌 Overview

Pakistan Law Chatbot is a Retrieval-Augmented Generation application designed to answer questions using information contained in uploaded Pakistani law documents.

The system combines:

• PDF document processing
• text chunking
• semantic embeddings
• vector retrieval
• local LLM inference
• conversational interaction

Instead of generating an answer from the language model alone, the application first retrieves relevant content from the uploaded documents and provides that context to the model.

---

🎯 Problem

Legal documents can be long and difficult to search manually.

This project explores how a RAG pipeline can make document-based legal information easier to access through natural-language questions.

The application is designed for document exploration and information retrieval, not as a replacement for a qualified legal professional.

---

✨ Features

• 📄 Multi-PDF document upload
• 🔍 Automatic PDF text extraction
• ✂️ Configurable text chunking
• 🧠 Semantic search using Hugging Face embeddings
• 🗄️ Persistent local ChromaDB vector storage
• 🤖 Local LLM inference through Ollama
• 💬 Natural-language question answering
• 🖥️ Simple Gradio interface
• 🏠 Local-first processing workflow

---

🏗️ RAG Architecture

               PDF Documents
                     │
                     ▼
              Text Extraction
                     │
                     ▼
                 Chunking
                     │
                     ▼
        Hugging Face Embeddings
                     │
                     ▼
               ChromaDB
            Vector Database
                     │
                     │
User Question ──────┘
        │
        ▼
 Semantic Retrieval / Top-K
        │
        ▼
 Retrieved Document Context
        │
        ▼
 Ollama Local LLM
        │
        ▼
 Generated Answer

---

🧠 How It Works

1. Document Ingestion

Users upload one or more Pakistani law PDFs.

2. Text Extraction

Text is extracted from the uploaded PDF documents.

3. Chunking

Large documents are divided into smaller chunks to improve retrieval.

4. Embedding Generation

The text chunks are converted into vector representations using a Hugging Face embedding model.

5. Vector Storage

Generated embeddings are stored in a persistent ChromaDB collection.

6. Semantic Retrieval

When the user asks a question, the system searches for the most relevant document chunks.

7. Answer Generation

The retrieved context is provided to the locally running Ollama model to generate the response.

---

🛠️ Tech Stack

Component	Technology
Language	Python
RAG Framework	LangChain
Embeddings	Hugging Face / Sentence Transformers
Vector Database	ChromaDB
LLM Runtime	Ollama
LLM	LLaMA-family model
Interface	Gradio
Document Processing	PyPDF

---

📂 Project Structure

Pakistan-Law-Chatbot-RAG-System/
│
├── Law-chatbot.py
│   └── Document ingestion, retrieval,
│       generation, and Gradio interface
│
└── README.md

The current implementation is intentionally compact, with the main application logic contained in Law-chatbot.py.

---

⚙️ Getting Started

Prerequisites

Install:

• Python 3.x
• Git
• Ollama

---

1. Clone the Repository

git clone https://github.com/abdullahk970/Pakistan-Law-Chatbot-RAG-System.git

cd Pakistan-Law-Chatbot-RAG-System

---

2. Create a Virtual Environment

python -m venv venv

Windows

venv\Scripts\activate

Linux / macOS

source venv/bin/activate

---

3. Install Dependencies

pip install langchain chromadb sentence-transformers gradio pypdf

---

4. Start Ollama

Start the local Ollama service:

ollama serve

Then download the model used by the application:

ollama pull llama2

Make sure the model name matches the configuration used by the current application.

---

▶️ Run the Application

python Law-chatbot.py

The Gradio interface will provide a local URL for interacting with the chatbot.

---

📡 RAG Pipeline

The core retrieval flow can be summarized as:

Documents
   ↓
Text Extraction
   ↓
Chunking
   ↓
Embeddings
   ↓
ChromaDB
   ↓
Top-K Retrieval
   ↓
Retrieved Context
   ↓
Local LLM
   ↓
Answer

---

🧪 Evaluation

RAG quality should be evaluated using a defined set of questions and expected source documents.

Useful evaluation metrics for this project include:

• Retrieval Recall@K
• Retrieval Precision@K
• MRR
• Answer relevance
• Context relevance
• Faithfulness of generated answers

No numerical benchmark is claimed here until a reproducible evaluation dataset and methodology are defined.

---

⚠️ Limitations

This project has several limitations:

• Retrieval quality depends on document chunking and embedding quality.
• The language model can generate incorrect or incomplete responses.
• Similarity retrieval does not guarantee that the retrieved passage is legally sufficient.
• Uploaded PDFs may contain formatting or extraction problems.
• The current implementation does not provide formal legal verification.
• Answers should not be treated as professional legal advice.

---

🔐 Privacy & Security Considerations

Because the application may process legal documents, production deployments should carefully consider:

• secure document handling
• temporary-file cleanup
• protection of stored vector data
• access control
• avoiding sensitive information in logs
• secure model/service configuration

Do not upload confidential legal documents to an environment that is not trusted or properly secured.

---

🔮 Future Improvements

Potential improvements include:

• Urdu-language query support
• Citation-based answers
• Page and section references
• Hybrid retrieval
• Retrieval reranking
• RAG evaluation dataset
• Better handling of scanned PDFs
• Document metadata filtering
• Improved answer verification
• Web deployment with appropriate security controls

---

👨‍💻 Author

Muhammad Abdullah Khan

• GitHub: abdullahk970
• LinkedIn: Muhammad Abdullah Khan

---

📄 License

This project is licensed under the MIT License.