Mullion RAG Demo

Role-based access control with fork/merge patterns for sensitive data

The Scenario

A Retrieval-Augmented Generation (RAG) system provides answers based on a document corpus with varying access levels. Documents are classified as Public, Internal, or Confidential. The challenge: ensure users only receive answers based on documents they have clearance to access.

The Problem

Traditional RAG systems risk exposing sensitive information through:

Accidental retrieval of confidential documents
Cross-contamination between security contexts
Lack of provenance tracking for sources

Mullion's Solution

Mullion enforces access control at the type level using:

Scope-based filtering: Documents tagged with access level scopes
Fork/merge patterns: Parallel processing with cache optimization
Provenance tracking: Know exactly which documents influenced the answer

How It Works

Query Classification

Analyze user query to understand intent and scope

↓

Access-Controlled Retrieval

Filter documents based on user's role (Public/Internal/Confidential)

↓

Parallel Processing (Fork)

Process retrieved documents in parallel with cache optimization

↓

Response Generation (Merge)

Combine insights while maintaining scope boundaries and provenance

✅ RAG with Mullion

rag-pipeline.ts

// ✅ SAFE: Mullion enforces access control
const userRole = 'internal'; // Public, Internal, or Confidential
const ragCtx = scope(`rag-${userRole}`);

// Query classification
const query = await ragCtx.infer(QuerySchema, userInput);

// Retrieve only accessible documents
const docs = await retrieveDocuments(userRole, query.value);

// Fork/merge for parallel processing
const result = await fork({
  branches: docs.map(doc => ({
    fn: () => ragCtx.infer(AnswerSchema, `...${doc}...`)
  })),
  strategy: 'cache-optimized',
  merge: weightedVote()
});

// Response includes provenance
console.log(result.value); // Answer
console.log(result.provenance); // Source documents

Key Features

Access Control

Role-based document filtering enforced at compile-time

Fork/Merge

Parallel processing with automatic cache optimization

Smart Caching

Provider-aware caching reduces costs and latency

Provenance

Track which documents influenced each answer

Try It Yourself

Query a document corpus with different access levels and see how Mullion enforces role-based boundaries.