Michael Rowe

Knowledge graph

note6 min read

In a knowledge graph, the connections carry the meaning

We understand concepts through their connections to other concepts, theories through their critiques and extensions, methods through the traditions they came out of. A knowledge graph makes that relational structure explicit and computable, treating the connections between ideas as carrying as much information as the ideas themselves.

Knowledge graph

One-sentence definition: A structured representation of knowledge that connects entities through explicit, typed relationships, enabling traversal1 and reasoning over conceptual connections rather than statistical similarity.

Scholars have always built knowledge graphs, even if we didn’t call them that. The citation network showing who built on whose work, the conceptual map linking related theories, the methodological genealogy tracing an intellectual tradition — each is a network of entities connected by meaningful relationships.

What’s changed is that these structures can now be made computationally explicit, so a system can reason about them. Not simply that Paper A and Paper B both discuss validity, but that Paper A critiques Paper B’s approach to validity, that both draw on Methodology X, and that the same critique reaches Framework Y.

How they work

A knowledge graph consists of nodes and edges. Nodes represent the entities you want to reason about — concepts, people, papers, methods, theories — and carry properties of their own: a scholar has an affiliation and research interests, a paper has a publication year and a methodology, a concept has a definition and a domain.

Edges represent typed relationships, which say how two things are connected and in which direction: Schön built his account of reflective practice on Dewey’s account of reflective thinking, and Paper A extends Paper B while challenging some of its assumptions. Properties can attach to edges as well as nodes, so a ‘critiques’ relationship can record the nature of the critique, the year it was published, and whether the person critiqued ever acknowledged it.

Written out, the structure looks like this: (Paper A) --[extends while challenging]--> (Paper B)

A programme’s curriculum map is a knowledge graph most educators already maintain, usually as a spreadsheet. The nodes are professional standards, learning outcomes, modules, assessments, and placements; the edges say which outcome maps to which standard, which assessment tests which outcome, and which module prepares students for which placement. Held as a spreadsheet, the map answers questions one row at a time. Held as a graph, it can answer questions that follow the edges, such as which of the regulator’s standards are only ever assessed on placement, or which outcomes are tested by an assessment that no earlier module prepares students for.

Knowledge graph vs vector database

Vector databases store numerical representations of text, called embeddings, that support similarity search. That works for finding passages about similar topics, but it can’t answer questions about how concepts relate, because embeddings capture co-occurrence patterns rather than explicit relationships.

The difference lies in what each treats as meaning. Embeddings encode the linguist J. R. Firth’s principle that you know a word by the company it keeps (Firth, 1957), which is distributional semantics2 built on statistical patterns. Knowledge graphs encode a different principle, that you know a concept by its connections, which is structural semantics3 built on relationships someone has stated.

Query a vector database for ‘social constructivism’ and you’ll find the passages that discuss it most. Query a knowledge graph and you can ask which critiques of social constructivism also apply to phenomenology, a question about shared methodological foundations that no amount of similar text will answer. This is why knowledge graphs support multi-hop reasoning while vector databases support only single-hop retrieval.

What becomes possible once the graph is explicit

One use is following chains of critique and influence through a citation network, which the multi-hop reasoning note works through. Two more are specific to holding your own field as a graph.

Methodological genealogy maps where methods come from, what assumptions they carry, and how they’ve changed, which needs explicit relationships between methods, traditions, and underlying commitments.

Research trajectory mapping shows how your own thinking has developed: which ideas led to which, which readings shaped which arguments, which collaborations produced which projects.

In context engineering, the graph is what lets AI reason about how your scholarship fits together. The linked notes in Obsidian, the conceptual maps, and the citation networks you trace are already knowledge graphs; the question is whether to make them explicit enough to traverse.

The curation challenge

Building a knowledge graph takes work. Automated extraction through GraphRAG helps, but it makes errors, and scholarly nuance is easily lost in the process — subtle distinctions collapsed, contested relationships stated as settled, contextual qualifications stripped out.

So a graph needs active curation. The relationship between two scholars is rarely just ‘cites’ or ‘critiques’; it might be ‘applies to a new context’ or ‘accepts the method but rejects the conclusions’, and characterising it accurately takes scholarly judgement.

The investment makes sense when you’re building knowledge infrastructure you’ll return to, such as a research area you’re committed to or a personal knowledge base supporting ongoing work. For a one-off literature review, reading will do. For building systems that can reason about your field, the explicit graph is what makes it possible.

What remains difficult

How to represent contested or ambiguous relationships, when scholarly disagreement is productive rather than a fault to be corrected. What granularity is right: concepts, arguments, papers, authors. How to handle understanding that shifts as your reading deepens. When explicit structure earns the effort, and when trusting implicit knowledge is enough.

Knowledge graphs work best for stabilised knowledge where relationships are broadly agreed. They struggle in emerging fields where every claim is contested, and in highly interpretive domains where relationships resist formalisation. The useful question isn’t whether a graph beats other representations but when its benefits — structured reasoning, explicit connections, knowledge that others can build on — outweigh what it costs to curate and maintain.

Sources

Notes

Most context engineering uses both representations, with retrieval over embeddings to find material and a graph to reason about how it connects.

Footnotes

  1. Traversal: moving through a graph by following its edges from one node to the next, the way you’d trace a learning outcome to the module that teaches it and then to the assessment that tests it. Wikipedia ↩

  2. Distributional semantics: the idea that a word’s meaning can be inferred from the words it tends to appear alongside. It’s the principle behind embeddings: two passages end up close together because they use similar words in similar ways, whether or not they make the same claim. Wikipedia ↩

  3. Structural semantics: here, meaning carried by explicitly stated relationships between concepts, as in a semantic network, where ‘assessed by’ or ‘prerequisite for’ is written into the structure instead of inferred from word patterns. Wikipedia: semantic network ↩

Continue reading

  • Context engineering

    How context engineering differs from prompt engineering: building the knowledge and structure a language model needs to reason well about your work.

  • GraphRAG

    GraphRAG combines knowledge graphs with retrieval-augmented generation. Where ordinary RAG returns the passages most similar to a question, GraphRAG extracts the entities and relationships in a document collection and answers by following the connections between them, which is what questions about critiques, influences, or curriculum coverage need.

  • Multi-hop reasoning

    Multi-hop reasoning with knowledge graphs: how AI answers questions that take several steps, and why recorded relationships make each step checkable.