THE AI-NATIVE KNOWLEDGE · GraphRAG: Designing Knowledge Retrieval on Graphs — From Core Concepts to Twelve Real-World Use Cases

Your search box can find documents. It cannot answer questions.

Somewhere in your organization there is a question no single passage contains: “Which form does this process need, and who signs it?” — “How many products share this component?” — “What changed since yesterday?” Classic RAG fails on all of them — and it fails quietly, with answers that sound confident and are wrong.

GraphRAG is the engineering discipline that fixes this — and this book teaches it end to end: not as theory, but as a complete, measurable method you can defend in front of your team.

What you will be able to do after reading:

• Decide with evidence, not fashion — classify your real questions, walk a 35-point decision matrix, and know exactly when you need a graph (and when you don’t)


• Build the graph from any source — LLM extraction, rules, database import, and images; entity resolution that never merges the wrong people


• Master the five retrieval patterns — local, global, DRIFT, path traversal, and Text2Cypher — and route every question to the right one


• Ship answers people can trust — citations that open, reasoning paths that are recorded (never invented), and refusals that are honest


• Evaluate and operate for real — golden sets, three-tier diagnosis, incremental updates, cost budgets, and access control that never leaks


• Go agentic when it pays — multi-step research loops with tools, budgets, and guards

What makes this book different:

• Three complete projects built chapter by chapter on realistic synthetic data — an internal knowledge assistant, a support chatbot, and a research assistant — with real token bills, real failures, and real fixes


• Nine industry deep-dives: code assistants, legal contracts, BI, CRM, e-commerce, medicine, fraud detection, news intelligence, and technical manuals


• Runnable companion datasets with 45 expert-keyed test questions and planted traps — so you can reproduce every number in the book


• Self-check quizzes in every chapter, with answer keys

Written for engineers, architects, and technical leaders building AI systems over private data. Python examples throughout; works with NetworkX for learning and Neo4j for production.

Measure first. Decide with criteria. Cite everything. Refuse honestly. That is GraphRAG as this book teaches it — 31 chapters, one method, and a system you can keep honest for years.