Skip to content
Multigrid

Knowledge graphs and structured knowledge

How to model, extract, resolve and query structured knowledge — with the queries that similarity search cannot answer, and the parts that break at scale.

A knowledge graph is not a technology choice. It is the decision to write down the entities in your business and the relationships between them, with stable identities, so that a question about two things at once has an answer. Most organisations already hold that information; it is spread over three CRMs, a spreadsheet, and the part of the wiki nobody has read since 2021.

These pages take the engineering path through it. What a triple is and what it costs you. How to get entities out of text without the model inventing relations. How to decide that two records are the same thing, with a threshold you can defend. How to give every fact a source and a date. And the queries — multi-hop, aggregate, negation, completeness — that a vector index structurally cannot answer, each written out in SPARQL, Cypher or SQL so you can run it rather than take it on trust.

Where this subject usually goes wrong is at the two ends: too much ontology before there is any data, or an LLM pointed at a corpus with no schema and no validation. Both produce a graph nobody queries. The middle path is narrow and it is what these pages describe.

What a Knowledge Graph Is, Concretely

A knowledge graph built from one ordinary sentence, triple by triple, with the same facts shown as tables and as a graph so the difference is visible.

8 min read

Ontologies, Taxonomies and Schemas

Three levels of structure — controlled vocabulary, taxonomy, ontology — with what each one buys you and what it costs to maintain.

9 min read

RDF, SPARQL and the Semantic Web, Twenty Years On

What RDF and SPARQL actually are, written as a working query language guide, plus a straight account of which parts of the semantic web vision were adopted.

10 min read

Property Graphs and Cypher: Model, Load and Query in One Evening

A complete first property graph: constraints, a CSV load, the queries worth knowing, and the MERGE mistake that duplicates every node.

11 min read

Extracting a Knowledge Graph With an LLM

A closed schema, a prompt that demands evidence spans, and a validation pass that rejects every entity and relation not grounded in the source text.

12 min read

Entity Resolution: Deciding Two Records Are One Thing

Blocking, pairwise scoring, and a threshold chosen from a precision/recall table with the asymmetric cost of a bad merge worked out on the page.

13 min read

Canonical IDs, and Why Your Graph Forks Without Them

An identifier design that survives merges, splits and re-ingestion: opaque internal ids, an alias table, tombstones, and ids that never stop resolving.

12 min read

Temporal Knowledge: Facts That Were True Then

Valid time and transaction time on edges, the three ways to store an interval, and the as-of and overlap queries the intervals make possible.

11 min read

Provenance in a Knowledge Base

How to attach a source to every fact — named graphs, RDF-star, edge properties or an assertion table — and the retraction query that makes it worth the storage.

10 min read

When to Use a Knowledge Graph: Questions Vectors Cannot Answer

Five classes of question that similarity search structurally cannot answer — multi-hop, aggregation, negation, conjunctive constraints and completeness — each with the query that does.

15 min read

Text-to-Cypher and Text-to-SPARQL

Serialising the schema into the prompt, validating the generated query against it, and executing read-only with a timeout — in that order, never the reverse.

12 min read

Keeping a Knowledge Graph Fresh

Change data capture into a graph: idempotent upserts, out-of-order events, deletes that must not delete, and a conflict rule written down before it is needed.

12 min read

Wikidata as a Free Backbone

How Wikidata is structured, how to query it properly with p/ps/pq rather than wdt, how to link your entities to it, and how to measure its coverage of your own domain.

13 min read

Product Catalogues as Graphs

Modelling products, variants, typed attributes and compatibility once, so that search facets, fitment questions and recall impact all fall out of the same graph.

11 min read

A Rules Engine Next to a Model

Which decisions must never be probabilistic, why a rules engine gives three things a model cannot, and how to route a request between them.

11 min read

Constraint Solvers and the Problems They Own

A rostering problem with its search space computed on the page, the CP-SAT model that solves it, and what a solver certifies that a sampler cannot.

12 min read

Checking Model Output Against a Database

A grounding check that requires every factual claim to be typed, looks each one up, and returns a citation, a correction or a refusal — never an unchecked assertion.

12 min read

A Glossary Your Company and Your Model Both Use

One term table, three consumers: selective prompt injection, query expansion in retrieval, and a validator that flags deprecated terminology in output.

10 min read

Master Data Management for AI Teams

The discipline behind deciding which customer record is the real one, in the four architectural styles, and why AI teams meet it whether or not they wanted to.

11 min read

Turning a Wiki Into Something Machines Can Query

Triage the pages, take the tables and infoboxes first, extract the rest against a schema, and work out whether the review is affordable before you start.

12 min read

Other topics