🛡️ GuardRAG Enterprise

Privacy-First, 100% Offline AI Document Intelligence & Retrieval

Powered by Local LLMs, 4-Tier Safety Guardrails, Real-Time Token Streaming & Host-Controlled Multi-Device Sharing

Interact with sensitive enterprise documents with complete privacy and zero data leakage.
GuardRAG runs 100% locally on your hardware. No prompts, documents, or embeddings ever leave your device or internal network.
Every AI answer is strictly grounded in verifiable source citations and protected by multi-tier PII scrubbing, credential blocking, anti-jailbreak defenses, and customizable reasoning personas.

🌟 Key Features & Updates

GuardRAG is tailored for privacy-sensitive industries—Legal, Healthcare, Finance, and Enterprise IT—where data compliance and zero external telemetry are mandatory.

🏗️ Architecture & Security Pipeline

GuardRAG guarantees complete data isolation across ingestion, retrieval, and LLM generation:

graph TD
    User([User / Web UI / CLI]) -->|Raw Document| Ingest[Document Parser & Cleaner]
    Ingest -->|Sanitize Injection| SafetyCleaner[Indirect Injection Sanitizer]
    SafetyCleaner -->|Text Chunks| PIIRedactor[PII & Credential Redactor]
    PIIRedactor -->|Clean Chunks| Embedder[FastEmbed ONNX Vectorizer]
    Embedder --> VectorDB[(Local Vector Store: FAISS / Qdrant)]
    
    User -->|Prompt Query| InputGuard[Input Safety Guardrail Engine]
    InputGuard -->|Sanitized Query| Retriever[Semantic Context Retriever]
    Retriever --> VectorDB
    VectorDB -->|Top-K Grounded Chunks| PromptEngine[Persona & Reasoning Prompt Engine]
    PromptEngine -->|Context + Query| LocalLLM[Local LLM via Ollama / Cloud API]
    LocalLLM -->|Live Token Stream (SSE)| OutputGuard[Output Safety & Redaction Filter]
    OutputGuard -->|Verified & Grounded Stream| User

🛡️ 4-Tier Safety Guardrails

Configure your security baseline per-session or when generating shared LAN links:

🎭 Reasoning Profiles & Custom Personas

Tailor how GuardRAG processes and responds to document queries:

1. Balanced (Default): General-purpose reasoning delivering helpful, concise, and grounded answers.
2. Strict Privacy: Extra-cautious persona with aggressive redaction of sensitive contextual references.
3. Fast Summarizer: Direct, high-level executive bullet summaries for rapid document review.
4. Custom Profile Builder:
   • Define custom system persona instructions.
   • Configure chunk sizes (e.g. 500 – 4000 chars) and overlap ratios.
   • Specify custom blocked keywords and topics.
   • Settings persist automatically across sessions in the persistent storage directory.

🌐 Secure Multi-Device LAN Sharing

Host GuardRAG on your workstation or local server and collaborate across your team without exposing data to the public internet:

• Host Authority Mode: Only the host machine can alter security policies, switch reasoning profiles, or ingest/delete document indexes.
• Isolated Guest Experience: Guests access a streamlined interface with contextual suggested queries tailored to the active document.
• Granular Share Policies: Choose a privacy level (Public, Internal, Confidential, Restricted) specific to the shared session.
• Local Network Broadcast: Displays local IP access URLs (http://192.168.x.x:8000) for frictionless LAN connectivity.

📥 Installation

1. Install via pip

pip install guard-rag

Or install from source:

git clone https://github.com/sowmiyan-s/GUARD-RAG.git
cd GUARD-RAG
pip install -e ".[dev]"

2. Prerequisites

• Python 3.9+
• Ollama: For running local LLMs (e.g. gemma3:1b, llama3.1, deepseek-r1, mistral). Download from ollama.com.
• Windows Users: Ensure the Microsoft Visual C++ Redistributable is installed.

🚀 Quick Start

1. Launch the Web Interface

Start the GuardRAG server and open the web dashboard:

guard-rag

Or run via Python module:

python -m guardrag.cli.main

The server automatically displays local and LAN access URLs:

============================================================
  🛡️  GuardRAG Server Running
  Local URL:  http://localhost:8000
  LAN URL:    http://192.168.1.100:8000
============================================================

2. Docker & Container Deployment

Run GuardRAG with persistent volume storage and non-root execution:

# Using Docker Compose (Bundled GuardRAG + Ollama)
docker-compose up -d

# Or build and run standalone container
docker build -t guardrag:latest .
docker run -d -p 8000:8000 -v guardrag_data:/data guardrag:latest

3. Command Line Interface (CLI)

Run ad-hoc queries in headless environments, CI/CD scripts, or terminal workflows:

# Query a confidential contract with a local LLM
guard-rag --pdf path/to/contract.pdf --model llama3.1 --sensitivity Confidential

CLI Configuration Flags:

🐍 Python SDK Integration

Integrate GuardRAG into your custom Python services, agents, and data pipelines:

from guardrag import build_rag_chain, load_stored_rag_chain
from guardrag.utils.safety import check_input_safety, check_output_safety

# 1. Build and index document with PII and credential protection
db_id, chain = build_rag_chain(
    file_paths=["quarterly_financial_report.pdf"],
    model="llama3.1",
    sensitivity="Confidential",
    redact_pii=True,
    chunk_size=1000,
    chunk_overlap=200,
)

# 2. Query knowledge base safely
response = chain.invoke({
    "input": "Summarize the EBITDA performance and major risk factors.",
    "chat_history": []
})

print("AI Answer:\n", response["answer"])

# 3. Inspect grounded source citations
print("\n--- Verifiable Citations ---")
for doc in response.get("context", []):
    print(f"File: {doc.metadata.get('source')} | Page: {doc.metadata.get('page', 1)}")

📁 Supported Document Formats

GuardRAG parses, chunks, and indexes a wide range of document types:

• Documents: .pdf, .docx, .doc, .txt, .md
• Structured Data & Logs: .csv, .json, .log
• Source Code: .py

🧪 Testing & Validation

GuardRAG includes an automated test suite covering safety guardrails, PII redactions, multi-device isolation, background process handling, and API endpoints:

python -m pytest tests/ -v

🤝 Contributing & License

Contributions are welcome! Please feel free to submit pull requests or file issues.

GuardRAG is open-source software licensed under the MIT License.