GraphRAG Fundamentals

What you'll learn

• ✓Recognise the three question types that make vector RAG fail and explain why
• ✓Build and walk a knowledge graph by hand, then extract entities and relations with an LLM
• ✓Configure and run Microsoft GraphRAG indexing and read its outputs
• ✓Use local and global search and answer multi-hop questions
• ✓Estimate the token cost of GraphRAG and decide when it is worth it

About this course

Vector RAG works until the question needs more than one chunk. Multi-hop questions, global summary questions and questions about relations all fail the same way: the right chunk is never retrieved, because the answer lives in the connections between documents. GraphRAG fixes that with a knowledge graph — and this course shows exactly how, on a small fictional corpus with planted facts so you can verify every answer by hand.

Session 1 builds a naive RAG pipeline from scratch and breaks it on purpose on three question types. Session 2 covers knowledge-graph fundamentals: nodes, edges, building a graph by hand with NetworkX, walking it, and LLM-based entity and relation extraction with a prompt you can read. Session 3 is Microsoft GraphRAG hands-on: installation, the settings file line by line, the indexing pipeline, and the entities, relationships and communities it produces. Session 4 covers querying: local versus global search and what each is for, multi-hop reasoning, community visualisation, token cost, and — importantly — when not to use GraphRAG.

Every session is a notebook shown on screen while the narration explains each cell; the notebooks work with OpenAI or with a local Ollama model. Cost is measured, not assumed.

The follow-up course, *GraphRAG in Production*, takes this to real public documents with evaluation, provenance and agents.

**Licensing.** Both corpora are free to use: the Nulmari documents are written for this course (which is also why no model has memorised them), and the *A Christmas Carol* excerpt is public domain. Microsoft GraphRAG is MIT. `Resources/LICENSES.md` sets out the details, including the licences of the optional local Ollama models.

Curriculum5 sections · 17 lectures · 1h 47m

Read alongside the course