Why Trustworthy AI Needs a Semantic Layer

That question opened Ramon’s session. Ramon is a Software Fellow at Progress, a professor at Barcelona Tech and a co-founder of Nuclia, and he has spent his career on one problem: getting reliable answers out of knowledge. His argument was simple. A confident sentence is not trust. Trust comes from the system behind the sentence, and that system starts with a semantic layer.

Why Isn’t LLM Confidence the Same as Trust?

LLMs are trained to predict the next token. They’re fluent, and fluency feels trustworthy. But a confidence score only tells you how likely the words are, given the context the model retrieved through RAG. It can’t tell you which sources support the answer, which business definition was applied, whether you’d get the same answer twice, or who validated it.

Ramon defined trust as three things a system has to deliver: validation against the source, reliable data and knowledge underneath, and reproducibility, meaning someone else can rerun the process and confirm the result.

What Does a Semantic Layer Add That an LLM Can’t?

LLMs are good at language. They aren’t good at your business rules or your entities. Ramon’s point is that you don’t choose between AI and structured knowledge. You combine them, and the enterprise semantic layer is where the structured side lives.

An ontology defines the business model. For example, a supplier provides a material, and that material is subject to a regulation. Taxonomies and thesauri connect the many words for one concept: PET, resin code 1 and polyethylene terephthalate are the same material, and a general-purpose LLM can’t reliably map your company’s private vocabulary. A knowledge graph then represents your actual data, the specific suppliers, materials and certificates, connected through the ontology’s relationships. The ontology can even infer new facts, such as which regulation a certificate covers.

How Does a Knowledge Graph Improve RAG?

Retrieval is where the knowledge graph earns its keep. Ramon described HippoRAG 2, a GraphRAG technique that extracts facts from documents as triples, builds a knowledge graph, and combines semantic search with personalized PageRank to find the entities and relationships most relevant to a query. In his team’s work, it delivered about eight points more accuracy than standard retrieval, with the biggest gains on domain-specific questions.

But GraphRAG is only as good as the graph. Who validates the facts that go in?

How Do You Validate Entity Extraction at Scale?

Ask an LLM to extract entities and it will return a confident list. Ramon’s team found that state-of-the-art models started missing entities once the number of entity classes grew from 20 to 40 or 60.

So validation comes from agreement between independent checks. Does the term read like a material in this sentence? Has the domain model seen examples like it? Does it match the terminology in the semantic layer? When the checks agree, the entity is accepted.

For relationships, he uses a simple trick. Asking an LLM to describe how two entities relate invites hallucination. So ask a yes-or-no question instead: does Acme Films hold certificate EU 2049? He put the reliability of that narrow question at 95% to 99%.

Run that inside an agentic loop that reads new information, proposes facts, validates them and learns from feedback, and the knowledge layer stays current as the business changes.

Why Does Traceability Matter for the Context Layer?

Every accepted fact needs a record: the document chunk that supports it, the ontology rule applied, which checks passed, the threshold used, and when the decision was made. Information that’s valid today may not be valid tomorrow.

This matters because of context poisoning. If one wrong fact enters the context an AI agent works from, it can steer every decision that follows. The system has to know what it doesn’t know, and send uncertain facts to human review. In Ramon’s words, losing possibly good knowledge is better than injecting bad knowledge.

What Does This Look Like in Practice?

His example was a research agency asking which water treatment performs better across regions. One lab report says activated carbon cut PFAS concentration by 72% after 30 days. Another reports a 61% decrease from a different pilot. The taxonomy links “reduction” and “decrease,” and each claim is validated against its source. But the system won’t call the results directly comparable, because the second report doesn’t specify the 30-day window. The answer comes back with that caution attached, which is exactly what a decision-maker needs to see.

The Takeaway

A generated sentence doesn’t earn trust. The system around it does, through validation, traceability and reproducibility. For enterprises building RAG, GraphRAG and agentic RAG, that system is the semantic layer: governed definitions, a validated knowledge graph, and a record of every decision. Progress is building toward this across the Progress Data Platform, with Progress Semaphore providing the governed taxonomies and ontologies that give AI its meaning.

For Ramon's full presentation and other topics discussed at the Progress Data Platform Summit, view the recording.