Module 2 · Retrieval that reasons · Lesson 7

Knowledge graphs and GraphRAG

When questions are about relationships and aggregates rather than passages, how a schema'd knowledge graph built from verified records differs from LLM-extracted GraphRAG, what each costs, and what the evidence says about when they're worth it.

65 min 1 video 3 exercises 6-question quiz Answers

Setup for the exercises:Install Claude Code Python with uv

By the end of this lesson you can

Classify questions as passage, multi-hop, relational aggregate or global sensemaking, and match each to a retrieval approach
Build a small knowledge graph from verified ingest records with entity resolution by stable keys, and query it
Explain how Microsoft's GraphRAG indexes and answers queries, and estimate its indexing cost against vector RAG
Assess GraphRAG claims critically using recent unbiased evaluations

Some questions are not about any passage. "Which of our live matters have a charge in favour of Northfield Bank?" "Which leases in the Riverside Court development still use version 2 of our service charge clause?" "Has this title been bought or mortgaged through us before?" A vector index returns passages that mention Northfield Bank; it can't count them, join them to matters or guarantee it found them all. These are questions about entities and relationships, and they are what knowledge graphs are for.

"GraphRAG" is used for two quite different things. One is a schema'd knowledge graph you build deterministically from structured records you already trust. The other is Microsoft's GraphRAG and its relatives, which use an LLM to extract a graph from unstructured text and summarise it. This lesson covers both, what each costs, and the evidence on when either beats the retrieval you already have.

Knowledge graphs and GraphRAG

Knowledge graphs and GraphRAG

Knowledge graphs and GraphRAG

Sign in to keep reading.

Sign in to keep reading.