Skip to content

Ch 14. Capstone — A Domain Knowledge Graph + GraphRAG Pipeline

What you'll build

  • An end-to-end pipeline tying together every piece from Parts 1–5
  • Ontology design → graph construction → GraphRAG indexing → routed retrieval → evaluation
  • A step-by-step checklist to run one domain of your own all the way through

Assumes

All of Parts 1–5. This chapter is integration, not new concepts.

1. Goal — one domain, all the way

We've learned the pieces separately. The capstone ties them to one domain. Small and concrete is better — an internal wiki, a set of product docs, 50 papers in a specific field. Ideally the very corpus where "vector RAG couldn't solve multi-hop/global questions."

2. The pipeline — 5 stages

Capstone — domain KG + GraphRAG pipeline, 5 stages Capstone — domain KG + GraphRAG pipeline, 5 stages

The whole picture follows this course directly.

① Ontology design      (Part 2)  competency questions → core concepts/relations
② Graph construction   (Part 3)  DB schema + LLM text extraction → Neo4j
③ GraphRAG indexing    (Part 4)  community detection → community summaries
④ Routed retrieval     (Part 4)  classify question → vector / local / global
⑤ Evaluation & ops     (Part 5)  retrieval/answer evaluation, incremental updates

3. Step-by-step checklist

① Ontology (Ch 4)

  • Wrote 10–20 competency questions (reflecting real traffic).
  • Settled on 5–10 core concepts and a relation-type list.
  • Didn't over-engineer — only what the questions demand.

② Graph construction (Ch 5–Ch 8)

  • If you have structured data, extracted the skeleton from the DB schema.
  • Extracted unstructured text with the LLM using the ontology as a guardrail.
  • Merged duplicate nodes with entity resolution (auto only the certain ones, review the rest).
  • Pinned timestamps onto changing facts (where needed).

③ Indexing (Ch 9)

  • Ran community detection.
  • Generated per-community summaries (essential if you have global questions).

④ Retrieval (Ch 10–Ch 11)

  • Added a router that classifies questions into factual/multihop/global.
  • Local search gives graph facts + source chunks together.
  • Factual questions go to cheap vector RAG.

⑤ Evaluation & ops (Ch 12–Ch 13)

  • Built an eval set across the retrieval and answer layers.
  • Compared against a vector-RAG baseline to confirm the questions where GraphRAG earns its keep.
  • Set up an incremental update path (not full re-indexing).

4. What you walk away with

Run this pipeline through once and you return to the opening question — "does this domain really need GraphRAG?" The answer is set by your corpus and question distribution. If it's mostly factual questions, vector RAG is enough; if multi-hop and global are central, the graph earns its keep. The capstone's real deliverable isn't just "a working GraphRAG" but a firsthand-validated judgment of when to reach for a graph.

5. Where to go next

  • Deeper hybrid — vector entry + graph expansion, refined within a single question.
  • Agent memory — Ch 8's time-aware graph as an agent's long-term memory.
  • Ontology evolution — a loop that grows the ontology each time a new question type appears.

Closing

If vector RAG knows only "what's similar," a knowledge graph knows "how things connect." GraphRAG makes that connection retrievable, and its value lives in multi-hop and global questions. But it's no free lunch — routing by question, managing cost, and validating with evaluation are all part of one set.

Related guides: Ontology & Knowledge Graphs · Complete RAG Guide · AI Eval