🛡️ PolicyWise — AI Legal Policy Assistant

An intelligent assistant that analyzes policy and legal clauses using RAG, a custom ML risk classifier, and LLM-powered explanations.

---

✨ What is PolicyWise?

PolicyWise is an AI-powered tool that helps Compliance and Legal teams quickly evaluate policy or legal clauses.

It combines:

• RAG (Retrieval-Augmented Generation) → Finds relevant text inside uploaded PDF policies
• Machine Learning Classifier → Predicts if a clause is COMPLIANT or RISKY
• LLM Explanation (OpenAI) → Gives clear explanations and safer rewrites

This makes PolicyWise a smart internal assistant for reviewing documents.

---

🚀 Features

🔍 1. Document Search (RAG)

Upload PDF policy documents.
PolicyWise will:

• Extract text
• Break it into chunks
• Create embeddings
• Use FAISS to retrieve the most relevant sections

🛡️ 2. Risk Classifier (ML Model)

A Logistic Regression + TF-IDF classifier trained by me.
It predicts:

• COMPLIANT
• RISKY

With a confidence score.

🤖 3. AI Explanation (LLM-Enhanced)

If an OpenAI key is provided, PolicyWise can:

• Explain why a clause is risky
• Highlight dangerous wording
• Suggest a safer rewrite
• Use RAG + ML to give better, more contextual answers

---

📁 Project Structure

policy-wise/
│
├── app.py                 # Main Streamlit application
├── train_model.py         # Training script for ML classifier
├── requirements.txt       # Project dependencies
├── README.md              # Documentation
│
├── policy_model.pkl       # (Optional) Saved ML classifier
├── policy_vectorizer.pkl  # (Optional) Saved TF-IDF vectorizer
│
├── assets/
│   └── policywise.png     # Screenshot for README
│
├── .streamlit/
│   └── config.toml        # Technical blue theme for UI
│
├── LICENSE
└── .gitignore             # Ignored files (venv, .env, cache, etc.)

---

🧱 Architecture

Here’s a compact high-level overview of how PolicyWise processes, analyzes, and evaluates policy text:

🧑‍💻 User (Streamlit UI) → 📄 PDF Processing (Extract + Chunk + Embed)
→ 🔍 FAISS Search (RAG) → 🛡️ ML Classifier (TF-IDF + LR) → 📤 Final Output

🛠️ Installation

1️⃣ Create a virtual environment

python -m venv venv
venv\Scripts\activate

2️⃣ Install dependencies

pip install -r requirements.txt

3️⃣ Train the ML model

python train_model.py

4️⃣ Run the Streamlit application

streamlit run app.py

---

🚀 HOW IT WORKS

---

1. 📄 Upload Policy PDFs

   • Extract text using PyPDF
   • Split into overlapping chunks
   • Create embeddings (OpenAI)
   • Store vectors in FAISS index

2. ✍️ Enter a Clause

   • Convert clause → embedding
   • Search FAISS for top-matching policy snippets (RAG)

3. 🛡️ ML Risk Classification

   • TF-IDF vectorizer transforms text
   • Logistic Regression predicts: → ✅ COMPLIANT → ❌ RISKY
   • Outputs label + confidence score

4. 🤖 LLM Review

   • Combine: user clause + retrieved policy snippets + ML output
   • AI generates:
     • Explanation of risk
     • Highlighted vague phrases
     • A safer rewritten version

5. 📤 Final Output

   • ML prediction
   • Relevant policy snippets (RAG)
   • LLM explanation + rewrite

---

🧰 TECH STACK

---

FRONTEND 🖥️

Technology	Purpose
🎨 Streamlit	UI & user interaction
🐍 Python	Core language

BACKEND / PROCESSING ⚙️

Technology	Purpose
📄 PyPDF	Extract text from PDFs
✂️ Custom Chunking	Split policy text into chunks
🧠 OpenAI Embeddings	Convert text into vectors
🗃️ FAISS Vector DB	Fast semantic search (RAG)
📊 Scikit-learn	ML toolkit
🧩 TF-IDF Vectorizer	Transform text for ML model
🛡️ Logistic Regression	Classify COMPLIANT / RISKY

AI LAYER 🤖

Technology	Purpose
🧠 OpenAI Chat Models	Explanation + safer rewrite
🔍 RAG Pipeline	Retrieve relevant policy snippets

UTILITIES 🔧

Technology	Purpose
🔑 Python-dotenv	Load environment variables
💾 Pickle	Save model & vectorizer
🔢 NumPy	Numerical operations

---

💡 Why PolicyWise Matters

• Helps Legal teams quickly evaluate compliance risks
• Reduces manual effort in reviewing internal policies
• Uses a hybrid AI system (RAG + ML + LLM), similar to real enterprise tools
• Demonstrates applied knowledge of NLP, vector search, and model pipelines

---

🚧 Future Enhancements

• Add LegalBERT for deeper clause understanding
• Add metadata-based RAG (policy titles, categories)
• Deploy with Docker / Streamlit Cloud / HuggingFace
• Add authentication for internal company use
• Add clause history + downloadable reports