guard-rag 1.2.6

Privacy-first, fully offline AI document assistant secured by tiered safety guardrails

🛡️ GuardRAG

Privacy-First, Fully Offline AI Document Assistant

Secured by a Tiered Safety Guardrails System
v1.2.4 — High-Performance, Secure, and Professional

Upload any document. Ask anything. Get answers — entirely on your local machine or using secure cloud endpoints.
Supports offline local models and OpenAI-compatible cloud APIs with fully customized security boundaries.

---

💡 What's New in v1.2.4

• 🌐 OpenAI-Compatible Cloud Integration: Easily switch to high-performance cloud endpoints (Groq, OpenRouter, OpenAI, Anthropic, Cohere) with full API key security.
• 🔄 Dynamic Model & Host Switching: Swap Ollama hosts or LLM models directly from the frontend UI without restarting the application or re-uploading documents.
• 🛡️ Tiered Safety Guardrails: Fully offline safety engine (Public, Internal, Confidential, Restricted) preventing jailbreaks, credential leaks, and PII leaks.
• 🦁 Brave Shields & CORS Fallback: Enhanced compatibility featuring automatic backend proxy health checks when direct local browser fetches are blocked by privacy shields.
• 📚 Document Library: Persisted document collections with a visual session management panel (rehydrate or delete stored FAISS indices).
• ⚡ Faster Indexing: Optimized FAISS batching and retrieval parameters.

---

💡 Use Cases

GuardRAG is designed for professionals and organizations that handle sensitive data and require the power of LLMs without compromising privacy.

• 🔒 Secure Document Analysis: Chat with confidential contracts, internal financial reports, or legal documents without uploading them to a cloud provider.
• 🏥 Healthcare & Privacy: Analyze medical records or research data locally, ensuring compliance with privacy standards like HIPAA or GDPR through integrated PII detection.
• 💻 Offline Research: Work on-the-go or in air-gapped environments. Once the models are downloaded, no internet connection is required.
• 🛠️ Developer Productivity: Quickly query local documentation or large codebases (via text/PDF) using a streamlined CLI or Web interface.

---

⚙️ Data Sensitivity Tiers

Protect your information using our built-in safety engine (processed entirely offline):

Level	Protection Scope
🟢 Public	Detects jailbreaks and basic prompt injections.
🔵 Internal	Adds detection for API keys, credentials, and tokens.
🟡 Confidential	Protects SSNs, emails, phone numbers, and credit card info.
🔴 Restricted	Strict protection for medical history, HIPAA/GDPR, and financial data.

---

📥 Installation

Install the package directly from PyPI:

# Recommended stable version
pip install guard-rag

Prerequisites

1. Ollama (Optional for local execution): Download and install from ollama.com.
2. Model: Pull a model to use locally (e.g., ollama pull gemma3:1b).
3. Windows Users: You must have the Microsoft Visual C++ Redistributable installed to run the AI engine.

---

🚀 Commands & Usage

GuardRAG provides a flexible CLI and a web interface to interact with your documents.

1. Launch the Web Interface

Simply run the command with no arguments to start the local server and open the UI in your browser:

guard-rag

2. Chat with a Document via CLI

Start an interactive chat session directly in your terminal:

guard-rag --pdf path/to/my_document.pdf

3. Advanced Configuration & Cloud Endpoints

Set environment variables for cloud LLMs (Groq, OpenRouter, OpenAI, etc.):

# Optional API Keys for cloud execution
export OPENAI_API_KEY="your-openai-api-key"
export GROQ_API_KEY="your-groq-api-key"
export OPENROUTER_API_KEY="your-openrouter-api-key"

Customize the model, server, and safety levels:

# Custom local model
guard-rag --pdf report.pdf --model llama3 --sensitivity Confidential --chunk-size 1000

# Cloud model via Groq API
guard-rag --pdf report.pdf --model llama-3.1-8b-instant --ollama-host https://api.groq.com/openai/v1 --sensitivity Restricted

---

📖 Available CLI Options

Argument	Description	Default
--pdf <file>	Path to the PDF document you want to analyze.	Required for CLI
--model <name>	The Ollama or OpenAI-compatible model name.	gemma3:1b
--ollama-host	The URL of Ollama or OpenAI-compatible endpoint.	http://localhost:11434
--sensitivity	Safety level: Public, Internal, Confidential, Restricted.	Internal
--chunk-size	Size of document chunks for processing.	1000
--chunk-overlap	Overlap between document chunks.	200
--no-guardrails	Disable all safety checks (not recommended).	False
--help	Show all available commands and flags.	-

---

Built with ❤️ by Sowmiyan S

GitHub · PyPI