guard-rag 1.3.0

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

🛡️ GuardRAG Enterprise

Privacy-First, Fully Offline AI Document Intelligence

Secured by a 4-Tier Safety Guardrails Engine

Empower your organization to interact with sensitive documents securely.
GuardRAG runs entirely on your local infrastructure. No data leaves your network, and every AI response is strictly grounded and cited.
Protected by enterprise-grade data redaction, credential blocking, and anti-jailbreak safeguards.

---

🌟 Executive Summary

GuardRAG is engineered for professionals and organizations handling confidential data—Legal, Healthcare, Finance, and Enterprise IT. It provides a highly secure Retrieval-Augmented Generation (RAG) pipeline that lets you query internal documents without exposing sensitive information to public APIs or unauthorized internal LLM prompts.

Capability	Description
🛡️ Offline Security	Operates 100% offline using local Ollama LLMs and ONNX-accelerated vector embeddings.
🔐 Data Redaction	Context-aware PII scrubbing automatically redacts SSNs, emails, and names prior to indexing.
📊 Verifiable Citations	The AI never hallucinates its sources. Every answer includes verifiable document citations.
🌐 Modern Web Console	Intuitive, responsive web UI for managing vector libraries, adjusting LLM models, and interacting with data.
🔌 Cloud Extensibility	Drop-in support for OpenAI, Anthropic, Cohere, Groq, and OpenRouter for hybrid deployments.

---

🏗️ Enterprise Architecture

GuardRAG’s architecture is built on a strict boundary between document processing and LLM interaction, ensuring that sensitive data is filtered before it reaches the language model.

graph TD
    User([User]) -->|Web UI / CLI| API(FastAPI Backend)
    API -->|Raw Document| PII[PII Redactor & Filter]
    PII -->|Scrubbed Text| Embed[FastEmbed ONNX]
    Embed --> VectorDB[(FAISS / Qdrant)]
    API -->|Query| SafetyInput[Input Guardrails]
    SafetyInput --> Retriever[Semantic Retriever]
    Retriever --> VectorDB
    VectorDB -->|Context| RAG[GuardRAG Chain]
    RAG -->|Local / Cloud LLM| Response[Raw Answer]
    Response --> SafetyOutput[Output Guardrails]
    SafetyOutput -->|Safe Answer| User

---

🛡️ 4-Tier Safety Guardrails

Configure your security baseline on a per-session basis. The guardrail engine runs locally and does not depend on external moderation APIs.

Tier	Icon	Protection Level
Public	🟢	Base protection against prompt injections, jailbreaks, and DAN-mode attacks.
Internal	🔵	Public + Blocks exposure of API keys, bearer tokens, passwords, and server credentials.
Confidential	🟡	Internal + Scrubs SSNs, emails, phone numbers, and credit card numbers.
Restricted	🔴	Confidential + Certified lock for Medical/Financial data, diagnoses, and salary info.

---

📥 Installation

Install the production-ready package directly from PyPI.

pip install guard-rag

System Requirements:

1. Python 3.9+
2. Ollama: Required for offline LLM support. Download from ollama.com.
3. Windows Users: Must install the Microsoft Visual C++ Redistributable.

---

🚀 Quick Start

1. Web Console (Recommended)

Launch the integrated Web UI. GuardRAG will automatically start a local server and open your default browser.

guard-rag

The web console supports LAN sharing. Just navigate to the ACCESS (LAN) URL displayed in your terminal from any device on your network.

2. Command Line Interface (CLI)

For headless server environments or automated pipelines, use the CLI directly:

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

CLI Configuration Flags:

Flag	Description	Default
--pdf <file>	Path to the target document (PDF, TXT, DOCX)	Required
--model <name>	LLM name (e.g., gemma3:1b, llama3.1)	gemma3:1b
--ollama-host <url>	LLM Endpoint (Local or Cloud API)	http://localhost:11434
--sensitivity <lvl>	Set guardrail tier (Public, Internal, Confidential, Restricted)	Internal
--chunk-size <int>	Text segment token limit	1000

---

🐍 Python SDK Integration

GuardRAG can be integrated seamlessly into your existing Python pipelines.

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

# 1. Initialize an offline RAG pipeline with PII Redaction
db_id, chain = build_rag_chain(
    file_paths=["financial_q3.pdf"],
    model="llama3.1",
    sensitivity="Restricted",
    redact_pii=True,
)

# 2. Query the Knowledge Base safely
response = chain.invoke({
    "input": "Summarize the Q3 earnings.",
    "chat_history": []
})

print(response["answer"])

# 3. Access Citations
for citation in response.get("context", []):
    print(f"Cited Source: {citation.metadata['source']}")

---

🤝 Support & Enterprise Licensing

GuardRAG is an open-source project under the MIT License. For enterprise deployment assistance, custom cloud integrations, or feature requests, please open an issue or reach out to the maintainers.

Built with ❤️ by Sowmiyan S

GitHub Repository · Issue Tracker