LlamaIndex TS Zero-Egress Sanitization: Pre-Prompt In-Memory Masking

import { Document, VectorStoreIndex, Settings } from 'llamaindex';
import { scrubText, unscrubText } from '@privacyscrubber/sdk';

// 1. Raw Sensitive Document Input
const rawDocuments = [
  new Document({
    text: 'Client Alice Walker (SSN: 987-65-4321, DOB: 05/12/1988) holds Portfolio #8841-A with Chase Wealth Management. Target asset allocation: 70% equities, 30% fixed income.'
  })
];

// Ephemeral Session Token Vault (Store securely or in Redis for multi-tenant RAG)
const tokenRegistry = new Map();

// 2. Pre-Ingestion Sanitization Step (<2.5ms per 50KB chunk)
const sanitizedDocuments = rawDocuments.map((doc, idx) => {
  const { sanitizedText, sessionMap } = scrubText(doc.text, {
    profile: 'Financial',
    detectSecrets: true
  });

  const docId = `doc_${idx}_${Date.now()}`;
  tokenRegistry.set(docId, sessionMap);

  return new Document({
    text: sanitizedText,
    id_: docId
  });
});

// 3. Build Vector Index (Vector Store only indexes tokenized text!)
const index = await VectorStoreIndex.fromDocuments(sanitizedDocuments);

// 4. Query with Tokenized Context
const queryEngine = index.asQueryEngine();
const response = await queryEngine.query({
  query: 'What is the asset allocation for client [NAME_1]?'
});

// 5. Restore Original Identifiers for Authorized User
const activeMap = tokenRegistry.get(sanitizedDocuments[0].id_) || {};
const cleartextAnswer = unscrubText(response.toString(), activeMap);

console.log('Sanitized Response in Vector Store:');
console.log(response.toString());

console.log('\nFinal Detokenized Result in Local Memory:');
console.log(cleartextAnswer);