The Semantic Layer, From Scratch

Building a federated knowledge graph over Aurora and Redshift with Stardog, and putting an AI agent on top

Vinodh Thiagarajan

About this book. Fifteen short chapters that take a database engineer from "what is a triple?" to a working AI agent answering questions live across Aurora PostgreSQL and Amazon Redshift through Stardog. Each chapter is a self-contained page with hands-on steps in Stardog Studio. Inspired by the AWS Machine Learning blog post on building a semantic layer with Stardog and Amazon Bedrock AgentCore.

The system you will build, one chapter at a time. The preface shows the full diagram.

Who this is for

You are an engineer who knows databases very well: SQL, joins, schemas, Postgres, Redshift, ETL. You have never worked with RDF, triples, SPARQL, OWL, ontologies, reasoning rules, SHACL, Stardog, Voicebox or MCP, and you would like to understand them properly rather than copy a demo. This book takes you from "what is a triple?" to a working AI agent that answers business questions live across two AWS databases, explaining every new idea in terms of the SQL you already know and saying plainly where the comparison breaks.

How to read it

Keep Stardog Studio open beside the book, connected to the database kit-c360. Keep DBeaver handy to check answers in SQL.
Do every green "Try it in Studio" box yourself. Practice steps use scratch graphs and clean up after themselves; chapters marked build step leave something permanent that later chapters rely on, so do them in order.
Ask Claude whenever something is unclear. After each build step, press its "Build step N done" button and paste it to Claude with your results, so we can confirm them before the next step builds on them.
Code blocks have a Copy button. Pages follow your system's light or dark setting.

Time to finish: about 13 hours in total: roughly 4.6 hours of reading and 8.1 hours hands-on in Studio and DBeaver. One or two chapters per sitting works well. The build chapters take longer because you create something real.

#	Chapter	Time
	Preface: What we are building The goal, the finished architecture, the build track, what's already done, and how to read the book.	~15 min 15 read
1	One question, five systems Why data fragments, what a semantic layer is (and isn't), the blog's three layers, and our Aurora/Redshift split.	~30 min 15 read · 15 hands-on
2	The triple: the atom of knowledge RDF's one shape, subject → predicate → object; a SQL row becomes triples; following arrows instead of joining.	~30 min 15 read · 15 hands-on
3	Names, values and files IRIs, literals and datatypes, prefixes, Turtle files, and named graphs as folders.	~45 min 15 read · 30 hands-on
4	SPARQL for people who know SQL Triple patterns, the shared variable as a join, FILTER, OPTIONAL, GROUP BY, and updates, side by side with SQL.	~45 min 15 read · 30 hands-on
5	The T-Box, part 1: classes (RDFS) Schema vs data, classes and subclasses, and the open-world assumption. You create the c360 classes.	~45 min 15 read · 30 hands-on	Build step 1
6	The T-Box, part 2: properties and the c360 ontology (OWL) Object vs datatype properties, domain and range, inverses, and loading the full kit ontology.	~60 min 18 read · 42 hands-on	Build step 2
7	Where the data lives: materialized vs virtual Copying triples in vs virtual graphs over SQL, data sources, and the trade-offs.	~30 min 15 read · 15 hands-on
8	Mappings: teaching Stardog to read your tables SMS2 mappings and IRI templates; you create the Aurora virtual graphs, safe and PII.	~45 min 15 read · 30 hands-on	Build step 3
9	Federation: where the magic happens The Redshift mapping, the IRI trick, what happens inside a federated query, and the top-spenders answer.	~60 min 18 read · 42 hands-on	Build step 4
10	Reasoning: teaching the graph to conclude Inference at query time, Stardog rules, and the kit's five rules, including "big spender".	~45 min 15 read · 30 hands-on	Build step 5
11	SHACL: when you do want rules to reject data Closed-world validation with shapes: the NOT NULL and CHECK constraints of the graph world.	~45 min 15 read · 30 hands-on
12	Governance: same question, different answers Users, roles and named-graph security; HR sees SSNs, Marketing doesn't.	~45 min 12 read · 33 hands-on	Build step 6
13	The Stardog toolbox Cloud portal vs endpoint, Studio, Designer, Explorer, Voicebox, the MCP server, and knowledge kits.	~30 min 15 read · 15 hands-on
14	The agent: from question to answer An OpenAI/OpenRouter model with a SPARQL tool, end to end, and how it maps to Bedrock AgentCore.	~60 min 20 read · 40 hands-on	Build step 7
15	Build your own Voicebox The Chapter 14 agent as a chat app: Stardog login, chat history, visible SPARQL, a model picker, and what each question costs.	~60 min 20 read · 40 hands-on	Build step 8
16	At scale: huge ontologies and answers you can trust Schema retrieval for ontologies too big for a prompt, and a source registry plus completeness checks so the agent says "partial" instead of a confident wrong total.	~75 min 28 read · 47 hands-on	Build step 9
	Glossary Every term in the book, A to Z, with the chapter that introduces it.	reference
	Key ideas at a glance The fifteen ideas worth remembering, each linked to where it is taught.	~5 min 5 read
	Corrections log What a reader flagged while working through the build steps, and what changed, chapter by chapter.	reference