Semantic Context Lab
What does a knowledge graph, and an ontology on top of it, add to an AI answering questions over company data? Three approaches, 300 questions, one synthetic company.
Runs at semantic.latentconjecture.com · checking…
Details
- Started
- 7 Oct 2026
- Questions
- 300, asked in six worlds nobody had looked at
- Approaches
- Text search, graph, graph plus ontology
- Tests
- 14 fixed before the run, Holm-adjusted
- Stack
- FastAPI, React, Neo4j 5
- Data
- Entirely synthetic
- Language
- Python
- Last commit
- 6 Nov 2026
ConjectureGiving an AI a knowledge graph, and then an ontology, gets it more of the evidence it needs than text search alone.
FindingHalf right. On questions that need several linked records, the graph found 22.1 points more of the needed evidence than text search. The ontology’s averages were higher still, but none of its tests passed.
The write-up
Most companies already have their data in tables. The question here is narrower than “graph or no graph”: if you put a knowledge graph over that data, and an ontology over the graph, does an AI answering questions about the company get better evidence to work from?
Everything happens inside a made-up multinational company. Every record, document and name is synthetic, generated by a program that also plants the answer key, which none of the approaches can see.
Three ways to look things up
- A, text search. Conventional retrieval over the company’s documents.
- B, the graph. A plus the links between records: this incident affects that game, which runs on this platform version.
- C, the ontology. B plus a shared dictionary of what things mean, so an incident type in one country matches the same idea in another country’s own codes and language.
All three get the same questions, the same answer model and the same evidence budget of 2,000 tokens. Answers are scored by fixed rules, with two AI judges as a second opinion.
What it found
Every rule was written down before the official run, which used six worlds nobody had looked at during development. Of the 14 comparisons fixed in advance, one passed: on multi-hop questions the graph found 22.1 percentage points more of the needed evidence than text search (95 percent interval 10.4 to 34.1, Holm-adjusted p 0.030).
The other 13 showed no detectable difference, and that includes every test of the ontology against the graph. The ontology’s averages were the highest of the three, but averages are not a test, so this run does not show that the ontology helps. Each comparison rests on 24 question groups and one small local answer model, so a difference has to be large to show up at all.
The public version
The live site is a read-only copy. It replays the recorded official run, so you can watch a question go through all three approaches and see which links the graph followed, without a model running anywhere. It sleeps when nobody is using it, so the first visit may show a short wake-up page.