🛡️ PromptGuard — AI Prompt Security Platform

Real-time AI prompt threat detection powered by RAG, Groq, and llama-3.3-70b-versatile.

---

📌 Overview

PromptGuard is a single-file AI security platform that scans user prompts in real-time and classifies them as SAFE, SUSPICIOUS, or DANGEROUS. It uses Retrieval-Augmented Generation (RAG) to ground every verdict in your own security policy — uploading your rules once lets the system cite the exact policy clause violated in each scan.

Built entirely in vanilla HTML/CSS/JS with zero dependencies, it deploys to Netlify in under 60 seconds.

---

✨ Features

Feature	Description
🔍 Real-time Prompt Scanning	Classifies any prompt instantly with threat type, severity, and confidence
📚 RAG-powered Policy Grounding	Embeds your security policy and retrieves relevant chunks via cosine similarity
🤖 llama-3.3-70b-versatile	State-of-the-art LLM analysis running on Groq's ultra-fast inference API
🧮 768-dim Vector Embeddings	Uses nomic-embed-text-v1_5 for high-quality semantic search
💬 Policy Q&A Chatbot	Ask natural language questions about your AI governance rules
📊 Scan History & Stats	Tracks all scans with verdicts, timestamps, and RAG grounding status
⚡ Zero Dependencies	Pure HTML/CSS/JS — no build step, no framework, no npm install

---

🚨 Threat Detection Categories

• Prompt Injection — Hidden instructions that attempt to override system rules
• Jailbreak — Role-play or fictional framing used to bypass safety guidelines
• Data Exfiltration — Requests to extract API keys, credentials, or context data
• Social Engineering — Emotional manipulation to extract sensitive information
• Policy Bypass — Attempts to circumvent your organization's AI usage policy
• PII Exposure — Prompts containing or requesting personally identifiable information

---

🏗️ Architecture

User Prompt
     │
     ▼
┌─────────────────────────────────────┐
│         RAG Retrieval Layer         │
│  Query → nomic-embed-text (768-dim) │
│  → Cosine Similarity → Top-K Chunks │
└─────────────────────────────────────┘
     │  Relevant Policy Context
     ▼
┌─────────────────────────────────────┐
│      llama-3.3-70b-versatile        │
│         (via Groq API)              │
│  Structured JSON verdict output     │
└─────────────────────────────────────┘
     │
     ▼
  SAFE / SUSPICIOUS / DANGEROUS
  + Threat Type + Severity + Policy Citation

---

🚀 Getting Started

Prerequisites

• A free Groq API key — takes ~30 seconds to create

Setup

1. Clone the repository

git clone https://github.com/your-username/promptguard.git
cd promptguard

2. Add your Groq API key

Open index.html and replace the placeholder on this line:

const GROQ_KEY = "YOUR_GROQ_API_KEY_HERE";

3. Open in browser

# Just open the file — no server needed
open index.html

That's it. No npm install, no build step.

---

☁️ Deploy to Netlify

1. Push the repository to GitHub
2. Go to netlify.com → "Add new site" → "Import from Git"
3. Select your repo — Netlify auto-detects the HTML file
4. Click Deploy

Your site is live in under 60 seconds. Set the GROQ_KEY as a Netlify environment variable if you want to avoid committing it.

---

🗂️ Project Structure

promptguard/
│
├── index.html          # Entire application — UI, logic, RAG pipeline
└── README.md           # This file

---

📖 Usage Guide

1 — Scan a Prompt

Paste any AI prompt into the text area and click Analyze prompt. Use the sample chips to quickly load example attacks (injection, jailbreak, exfiltration, social engineering).

2 — Build a Knowledge Base (RAG)

Go to the Knowledge base tab. Choose a built-in policy template:

• OWASP LLM Top 10 — industry-standard LLM security rules
• Corporate AI Policy — enterprise AI usage guidelines
• Healthcare / HIPAA — PHI and compliance requirements

Or paste your own custom policy text. Click Build knowledge base — the system embeds every chunk into 768-dimensional vectors using nomic-embed-text-v1_5. Once built, all future scans are grounded in your policy.

3 — Policy Chat

Go to the Policy chat tab and ask natural language questions like:

• "What does our policy say about API keys in prompts?"
• "Which OWASP rule covers jailbreak attempts?"
• "What are the penalties for a data exfiltration attempt?"

Responses are semantically retrieved from your knowledge base and answered by the LLM.

4 — Review History

The History tab shows all past scans with verdicts, timestamps, and whether RAG grounding was active. Click any entry to re-analyze it.

---

🔧 Tech Stack

Layer	Technology
LLM	Meta llama-3.3-70b-versatile
Inference	Groq API
Embeddings	nomic-embed-text-v1_5 (768-dim)
Vector Search	In-browser cosine similarity
Frontend	Vanilla HTML / CSS / JavaScript
Deployment	Netlify

---

👥 Team

Name	Role	Handle
Mirza Yasir Abdullah Baig	Project Lead	@mirzayasirabdullahbaig
Sadia Usman	AI Engineer	@Frost_Flashlog3
Nazish Javeed	Backend Developer	@Nazishjaveed
Hira Mujeeb	UI/UX Developer	@Mystic_Tintp1t7
Hamna Munir	Security Analyst	@HamnaMunir

---

⚙️ Configuration

Variable	Location	Description
GROQ_KEY	index.html line ~180	Your Groq API key
BATCH	buildKB() function	Embedding batch size (default: 8)
k in retrieve()	analyze() / sendChat()	Number of retrieved chunks (default: 3–4)
maxTokens	groqChat()	Max LLM output tokens (default: 600)

---

🔒 Security Notes

• Never commit your Groq API key to a public repository. Use environment variables or Netlify's secret management.
• All vector search and chunking happens client-side in the browser — your policy text is never sent to any server other than Groq's embedding API.
• The app makes direct browser-to-Groq API calls. For production use, consider proxying through a backend to keep the API key server-side.

---

📄 License

MIT License — free to use, modify, and distribute.

---

🙏 Acknowledgements

• Groq — for blazing-fast LLM inference
• Meta AI — for the llama-3.3-70b-versatile model
• Nomic AI — for the nomic-embed-text embedding model
• OWASP — for the LLM Top 10 security framework