Module 19: Graph RAG, Caching and RAG Security

Module Overview

Production RAG is more than vector search. This module covers retrieval over knowledge graphs, the emerging "vectorless" pattern, caching for cost and latency wins, and the security layer required before any system touches user data.

Learning Objectives

• Describe Graph RAG over a knowledge graph such as Neo4j.
• Explain the vectorless (PageIndex) pattern and when it makes sense.
• Add semantic caching to a RAG pipeline and reason about hit rates.
• Apply PII masking, input/output guardrails, and prompt-injection defence to a RAG system.

Topics Covered

Graph RAG

• Building Graph RAG with graph databases like Neo4j
• When graphs win — multi-hop, relational, and entity-centric questions

Vectorless RAG

• PageIndex and document-tree retrieval

Caching in RAG

• Plain caching vs. semantic caching
• Cache invalidation in a retrieval system

Securing RAG

• PII masking
• Input and output guardrails
• Prompt-injection defence (direct and indirect)

Key Concepts & Terminology

Knowledge-graph retrieval, entity-relation triples, semantic cache hit, guardrails, PII redaction, indirect prompt injection.

Tools & Frameworks Referenced

Neo4j, PageIndex, semantic caching layers, Presidio (PII masking), NeMo Guardrails.

Prerequisites

Modules 16–18 (RAG foundations and scaling).