In this article, you will learn how to benchmark a deterministic 3-Tiered Graph-RAG system against a standard vector RAG pipeline on fact-dense queries, and what the results reveal about matching prompt complexity to model capacity.

Topics we will cover include:

How to build a synthetic, fact-dense sports dataset that intentionally introduces noisy, contradictory text into a vector database while keeping ground truth in a graph store.

How to implement and compare two retrieval functions — one for standard vector RAG and one for the 3-Tiered Graph-RAG — using ChromaDB and a local Flan-T5 model.

Why a small-capacity local LLM can underperform with complex, conflict-resolution prompts, and what this means for production RAG system design.

Introduction

In a previous article , a deterministic, 3-Tiered Graph-RAG system architecture was proposed to address “lossiness” in dense vector databases when dealing with atomic facts.…