Why Ontology and Structure Still Matter in the Age of LLMs

One thing I think gets underestimated in the current AI cycle is how much structured design still matters, even in systems that use large language models.

I’ve seen that very directly in work on automated tagging and retrieval-oriented AI pipelines.

It is tempting to imagine that an LLM alone can take raw text and infer everything you need. In practice, high-quality tagging often depends on much more than that. It depends on ontology design, candidate generation, alias handling, entity recognition, scoring logic, search infrastructure, and evaluation that reflects what “good” actually means for the task.

That is part of what made this work interesting to me.

The pipeline was not just “send text to a model.” It involved combining structured ontology design, spaCy-based entity recognition, candidate generation logic, LLM-based scoring through Bedrock, and OpenSearch-backed retrieval thinking into something that could be iterated on and improved.

That kind of work reinforces a lesson I keep coming back to: strong AI systems often come from the interaction between structure and intelligence, not from either one alone.

LLMs are powerful, but they become far more useful when they are embedded inside well-designed systems with clear representations, constraints, and evaluation loops.

That is one of the reasons I remain very interested in applied AI architecture, not just model capability.

---

Originally posted on LinkedIn.