Why My AI Strategies Will Use Knowledge Graphs in 2026

Perspective — Jan 12, 2025

By Piers Fawkes — 6 min read

For a while, I genuinely thought vector databases were going to solve everything. They were fast, flexible, and worked well with LLMs. But the limits of similarity without structure have become obvious.

I could embed documents, retrieve similar content, and feed it into a model with relatively little setup. For many early use cases, that was enough. And for a time, it worked. But as I started pushing the AI systems I was building beyond simple retrieval and summarization, the limits of vector databases became obvious.

Vectors are very good at similarity. They are much less good at structure.

When you rely on vectors alone, you’re effectively asking the model to infer relationships rather than reason over them. Concepts may be close or far in a latent space, but the system doesn’t know why. Hierarchies, causality, precedence, and dependency are all implicit.

That works, until it doesn’t. This is why knowledge graphs are quietly re-emerging in modern AI stacks. Not as a replacement for vectors, but as a complement to them.

Explicit Relationships

Knowledge graphs do something fundamentally different. They make relationships explicit. They define what connects to what, how strong those connections are, and in which direction they flow. Instead of leaving logic for the model to guess, they allow systems to follow it.

In practice, this matters most when AI is used for reasoning rather than retrieval. If you’re building an AI system that needs to respect domain rules, understand trade-offs, or explain its outputs, vectors alone will eventually plateau. By ‘domain rules’, I mean the constraints and priorities that experienced practitioners apply instinctively, even when they aren’t explicitly written down in the data.

Vectors can retrieve relevant information, but they don’t enforce coherence. They don’t encode judgment. And they don’t preserve the mental models experts actually use. Well-designed knowledge graphs do.

Correlation vs. Causation

They allow AI systems to distinguish between correlation and causation. They make it possible to weight signals differently. They let you encode domain-specific bias intentionally, rather than letting it emerge accidentally from training data.

None of this means vectors are obsolete. They’re extremely useful for discovery, retrieval, and flexibility. But on their own, they tend to converge toward generic answers, especially as domains become more complex.

The systems that perform best in practice combine both approaches: vectors for breadth and recall, and knowledge graphs for structure, reasoning, and control.

As AI moves from experimentation to infrastructure, the question changes. It’s no longer just “Can we find relevant information?” It becomes “Can we reason with that information in a way that reflects how the domain actually works?” That’s the gap knowledge graphs are filling again. Not because they’re new, but because the problems we’re trying to solve have matured.