Home Page
The Need for SENTRY
Login Page
Search Page
Upload Page
Pipeline
Demo

SENTRY

SENTRY (Secure ENgine for Trusted RAG Yield) makes sure your AI assistant only sees what you’re allowed to see — nothing more, nothing less — and always explains why.

Inspiration

With more companies using Retrieval-Augmented Generation (RAG) systems, we noticed a big worry: privacy.

RAG helps AI give more accurate, knowledge-based answers, but it can also:

Accidentally mix up data between tenants
Reveal sensitive information in responses
Be tricked by malicious prompt injections
Have embeddings manipulated by adversarial inputs

We asked: How can we make RAG safe, transparent, and trustworthy—without slowing it down?

That question led to SENTRY — a privacy-first gateway for RAG systems.

What it does

SENTRY makes sure your AI only sees what it’s allowed to see. It:

Blocks unsafe or malicious queries
Filters out sensitive data automatically
Enforces access rules by user or tenant
Keeps a full audit trail to show why every decision was made

In short: your AI stays smart, but now it’s also safe.

How we built it

We created SENTRY with a lightweight, modular design:

React frontend for easy interaction
FastAPI backend with SQLite for storage
FAISS + Sentence Transformers for fast vector search
Embedding scrambling and encryption for privacy
Git/GitHub for version control and collaboration

SENTRY works as a middleware that sits between your queries and your RAG system, enforcing rules at every step.

Challenges we ran into

Balancing privacy vs. answer quality: showing less context reduces exposure but can make answers less complete.
Keeping security lightweight: too many checks can slow down real-time responses.
Handling complex enterprise access rules without making the system hard to use.

Accomplishments we’re proud of

Built dual-layer gates to protect queries and retrieved data
Policy-driven filtering that works out of the box
Transparent logging so every action can be traced for compliance
Smooth integration without slowing down the RAG workflow

What we learned

Privacy matters as much as AI performance: protecting data is a core feature, not an afterthought
Trade-offs are inevitable: the more we protect, the more careful we must be about context
Lightweight security works best: it’s easy to add, and it doesn’t frustrate users