For the better part of ten years I analysed large knowledge graphs the slow way: incremental, manual deep dives with SPARQL queries and human intuition. So I was curious to see how a language model, connected to the data through Model Context Protocol servers, would approach the same task.
The subject was Italy’s national catalogue of semantic assets for public administrations, published at schema.gov.it. The experiment ran blind: the agent orchestrated the whole analysis itself.
First, the landscape
I started by asking about the overall structure. At the time of the analysis (June 2025), the agent reported:
- about 64,000 concepts in controlled vocabularies
- more than 35,000 municipalities with identifiers
- over 80 specialised ontologies, from livestock management to occupational safety
Then, interoperability between regions
I then pushed further: could it evaluate how well different regions could exchange data in specific sectors? The agent produced indicative scores:
- Social security and pensions: high, thanks to shared ontologies and standardised identifiers used across regions.
- Public services: high, built on a well-defined core public service vocabulary, common classifications and multilingual labels.
- Agriculture and livestock: good, with comprehensive shared ontologies but some regional variation in implementation.
These scores came from the agent’s own method, so they are a starting point for expert review.
What stood out
Italy has built one of the most sophisticated government semantic infrastructures I have seen. In agriculture, for example, some regions are close to plug-and-play compatible because they rest on the same ontological foundations.
What took me hours of exploration and domain expertise, the agent did in minutes through systematic querying and pattern recognition. It did not replace the expertise: I needed it to judge which findings were sound. But it changed the speed at which a landscape can be read.
This experiment was the starting point for measuring convergence in catalogues like this one. Cross-border analysis is the natural next step.