Although I agree with the separation being argued here I don’t find the practical implications that helpful especially in an age where LLMs will be writing a proportionally increasing amount of queries. In the same way a semantic layer can be inconsistent with itself, a semantic layer can be inconsistent with a written ontology that doesn’t technically tie to your actual system. What are the actual best ways that ontologies have been implemented? Ive heard of how it works at Palantir but are they the exception?