Graphs fast and slow

Opinion
Architecture
Kahneman’s System 1 and System 2 map onto knowledge graph and AI tooling: LLMs and vector search are fast, ontology reasoning and SHACL validation are slow.
Published

January 9, 2026

Illustration: Graphs fast and slow

I read Kahneman’s “Thinking, Fast and Slow” probably five times. It’s scientific entertainment, full of unexpected creative questions, much wisdom about human nature. Ideal for long trips and holidays. Accidentally or not, it describes how most “knowledge graph + AI” tooling doesn’t fail because a capability is missing but because nobody points out which cognitive system it’s supposed to be.

Let me recap the fast and slow. System 1 is fast, associative, pattern-completing. An LLM generating a Cypher or SPARQL query from a question. An extraction model proposing triple candidates from raw text, a vector index returning “similar enough.” It has no internal signal for when it’s confidently wrong. System 2 on the other hand is slow, rule-bound, deliberate. Ontology entailment, SHACL shape validation. It’s a reasoner checking whether a generated query or an extracted triple is even permitted by the schema. It doesn’t generate anything, it constrains and checks.

Almost every GraphRAG pipeline collapses this distinction in one of two directions. Many frameworks are all System 1, no checking layer at all. Generate the query, execute it, return whatever comes back with no step that asks whether the query was semantically sound in the first place. You have frameworks with System 2 in name only. SHACL bolted on for validation, running under open-world semantics which structurally cannot flag a hallucinated predicate as wrong. If the predicate isn’t asserted anywhere that’s “unknown” not “invalid.” The check looks rigorous and catches almost nothing it was built to catch.

Most people are not aware of their fast/slow thinking process. You need both. One for survival (fight or flight), one for planning/forecasting. The same is true for KG, you need an LPG (fast) and RDF (slow) system. The CyberBench paper is probably the best reference (paper, data and code). My main concern when explaining this to customers is the investment it represents in money/time/people. It’s huge. The grand solution they ask for exists, but it’s heavy and I am not aware of a plug-n-play platform. No, it’s not Palantir or Fluree or Foundry. They all take shortcuts. Maybe it’s inevitable from a business point of view, but unsatisfactory from an epistemic one.