GraphNews

This tutorial from Microsoft previews how to use an Azure function called azure_ai.extract() to process some natural language data, such as contracts or support tickets, and receive a knowledge graph at the end

Why ontologies, knowledge graphs, and a semantic orchestration layer are the missing pieces most AI architectures skip.

Everyone is building semantics for AI agents. I wonder if most are going to regret it. I’m observing two distinct conversations happening around semantics, and almost everyone is conflating them or worse, ignoring one entirely. - Semantics for AI agents: giving your agents the context they need to understand and operate on your data. Everyone’s talking about this. It’s hot, it’s urgent, it’s the easy sell. These are the rebranded “context layers” vendors, Databrick’s Genie Ontology, open context layer such as Google’s Open Knowledge Format, etc - Semantic interoperability: defining what your data actually means, consistently, across all your systems, so everything can align around a shared understanding. This has been a problem since the dawn of enterprise systems. This is why data catalogs emerged. Standards such as RDF and OWL have been designed for this. Lately, OSI is being developed for the BI/analytics use case. We lack incentives to address this because point to point solutions are quick and dirty solutions. Problem: most organizations building semantics today are doing it only for the AI use case. They want it fast. Quick and dirty. Let the model figure it out. Ship it. What’s going to happen? You’re going to end up with multiple agents, each carrying their own version of semantics, each giving you slightly different answers to the same question, brittle systems. Sound familiar? It’s the exact same story as your data dashboards, your data swamps. Just with a new layer on top. The Semantic Swamp. The organizations that will win, IMO, are the ones who recognize that the semantic work they’re doing for AI agents is (or should be) the same work they need for semantic interoperability across systems. The AI agent is just another consumer. A demanding one, because you’re dealing with unanticipated questions. Nobody’s asking a fixed query anymore. That’s exactly why the foundation matters more, not less. The Netflix example is a good one to look at. They’ve been investing in semantic foundations for years, long before the AI wave hit. Their problem: so many schemas, it’s not maintainable. They need a way to define once and reuse over. When agents came along, they weren’t scrambling. The infrastructure was there. It just worked. They are not outliers. See companies who are hiring ontologist, for example JPMC job posting for VP Ontologist: “Your role will involve utilizing ontologies and taxonomies to enhance data interoperability and management, preparing our data for AI applications.” Do the semantic work once. Define it properly. Map it across your systems. Then every new consumer, agents included, just plugs into that foundation instead of building its own version of reality. This is not an AI problem. It’s a leadership problem. And it has a solution if you’re willing to do the work. Check out links in comments Are you seeing organizations treat semantic work as foundations and interoperability, or just for AI? | 94 comments on LinkedIn

It's weekend, let's take a step back from tech and platforms. Ask a scientist "what's your ontology?" and you get a 50-year old fight about whether the wavefunction is real. Ask a knowledge graph engineer "what's your ontology?" and you get a class hierarchy with three subclasses and a cardinality constraint. Same word but wildly different perceptions. In science, "ontology" means: what does the math actually correspond to, out there, in reality? Is the quantum state the physical thing itself (wavefunction realism)? Or just an agent's belief about a sparser underlying reality? People have built entire careers, and several genuinely competing theories, purely on disagreeing about this. It's how science works. In knowledge graphs, we call a SKOS taxonomy or an OWL class tree an "ontology" and move on with our day. No one's losing sleep over whether ':hasPart' carves reality at the joints. The question is subtle but worth considering over a few cups of coffee: is this structure describing the world, or just organizing our data about it? A real ontology commits to what exists and why. Most of what we develop is a schema with a bit of delusion. It's useful, sure, but it rarely answers the question 'what is it actually doing?'. Maybe that's fine, maybe it isn't. I'm only challenging you a little. Every enterprise KG project eventually hits the same wall: two teams model "Customer," "Product," or "Risk" differently, and nobody can say which one is right because nobody was ever forced to commit to what those things actually are. You end up with a schema that's locally consistent and globally incoherent. It has perfectly valid SHACL shapes that quietly contradict each other the moment you try to federate across departments, or external partners. If your "Customer" entity is just a convenient bucket for whatever fields the CRM team needed this quarter, you'll ship fast and pay for it later in entity resolution hell, broken joins, and AI agents hallucinating relationships that were never really there. If you instead force an actual ontological commitment (what is a Customer, independent of which system is asking ) you get something that survives contact with a merger, a new product line, or a regulator asking "are these the same risk or two different ones?". Of course, I know I am advocating the long road here. It means thinking, confronting diverging opinions, changing points of view. This is where the business world can learn something from the scientific method. #KnowledgeGraphs #Ontology #Semantics

The graph database community has spent decades arguing about which side is right. On one side: Labeled Property Graphs (LPG).

Why Your LLM is Only as Smart as Your Data Contract

A working demo that puts business terms and dbt lineage into one open graph on Databricks, queried with openCypher in PuppyGraph, and grounds an AI agent's answer in that graph instead of a guess.

Databricks just put ontology at the centre of its biggest summit launch. Ali Ghodsi called Genie Ontology the secret sauce, a live graph of everything an organisation knows. That is the category winning, and there is a twist. The semantic web paper just turned twenty-five. It pitched ontologies for software agents. Those agents stayed specialist, while the technology went into production everywhere else, knowledge graphs and ontologies at Google, Netflix, the BBC, Siemens, AstraZeneca and the energy majors. Agents at scale, the use case the paper was really written for, came last. They are here now, and this week a 134 billion dollar company built its flagship agent on an ontology, the 2001 vision a quarter century late. 🧭 The one question the noise skips Not what it does, but whether you own it. 1️⃣ Can you move the meaning to another platform, or does it live inside theirs? 2️⃣ Can the system prove an action is allowed before an agent takes it, or only log what it did after? 3️⃣ Did you author the meaning, or does the platform learn and hold it? ⚖️ The verdict On all three, Genie Ontology is a learned context layer inside Databricks, open in the table-format sense, not the meaning sense. Their own architect conceded it: no OWL, no reasoner, no formal axioms. Ali Ghodsi himself scoped it as data grounding, meant to sit alongside agents like Claude Code, not the layer that decides what an agent may do. None of that makes it weak. Grounding in governed data rather than documents is the right instinct. The mistake is letting it become the place your meaning lives. Boards have paid for that once already, with ERP, when the logic that ran the business turned out to sit inside the vendor's model and could not be taken anywhere. Agents only make the bill larger, because now that meaning is what an autonomous system acts on. The model you can rent. The meaning you should not. | 50 comments on LinkedIn

“Context Graph” has quickly become one of the most overloaded terms in the 2026 AI stack. Ask five vendors what it means, and you’ll get five incompatible answers, none of […]

After years of working with ontologies, I've become less interested in debates about RDF vs. LPG databases because both solve real problems, both have limitations, and neither is the right answer for every use case. What I've noticed across the enterprise projects I've been involved in is that teams choose a property graph for perfectly valid reasons: - Simpler development experience - Better integration with existing systems - Operational familiarity - Performance characteristics that fit their workloads However, the challenge appears later: as the graph grows, the questions shift from data to meaning. - What does this relationship actually represent? - Which connections are valid? - How should a new entity be classified? - What business rules govern the graph? - What assumptions are embedded in the model? In an ontology-driven environment, these questions are addressed explicitly through classes, properties, axioms, constraints, and semantic definitions. In a property graph, however, they often end up scattered across documentation, application code, and undocumented institutional knowledge. This is where I've seen metagraphs become valuable... not as a replacement for ontologies (since they can work together), not as a substitute for semantic reasoning, but as a practical way to make the graph describe itself. The operational graph captures business facts, while the metagraph captures the structure, semantics, constraints, and modeling decisions behind those facts. Does it provide the same capabilities as an ontology managed in a triplestore? Of course, not. You don't automatically get formal semantics, logical inference, standards-based interoperability, or the maturity of the Semantic Web stack. However, in environments where a property graph is already the chosen platform, a metagraph can significantly reduce the gap between connected data and understood data. In addition, an interesting consequence of metagraphs is that they don't just help humans understand a graph... they can also help AI systems. As GraphRAG systems, graph-based agents, and knowledge-driven retrieval architectures continue to evolve, having an explicit representation of a graph's semantics may become as important as the graph data itself. That's a much deeper topic... one I'll save for another post 😊

Contribute to aws-samples/sample-dot-ld-knowledge-graph-syntax development by creating an account on GitHub.

What a typed property-graph schema actually buys you in production

Neo4j’s Create Context Graph ships with a universal starting schema. Here’s why starting with 5 node types beats starting from scratch.

Why AI-generated ontology proposals need human validation before generated structure enters the operating layer

I am trying to build a living memory/context engine for my business, something that can remember projects, decisions, timelines, risks, and...

Building an Ontology for a Knowledge Graph Is a Process, Not a Diagram Five established approaches, and the ongoing process that keeps them working. Why I wrote this My recent article on POLE+O — …

Not a fan of Knowledge Graphs, but recently I started using them for a surprising reason: to build non-trivial private verifiers for agentic search. For those who don't know, building a private eval set for a scaffolded LLM in 2026 is really challenging, like seriously hard (bc scaffolded LLM is overpowered). It takes a lot of effort to find a question that's non-trivial to a scaffolded LLM yet still answerable. To find those question-answer pairs, I built a knowledge graph extractor where you can throw a corpus at it, and it extracts the entity relations using Qwen3.6-35b-A3b-MTP on an L4 at 70 tps (which is really good for such a low-budget GPU). Then I mark out the longest path in the graph and use it to generate challenging question-answer pairs. The idea is to find those genuinely multi-hop fact chains that are verifiable from the corpus, to stress-test the agentic search system. | 70 comments on LinkedIn

After writing the previous article, What is a Financial Product at all?, I originally planned to move directly into a concrete overview example of ISDA CDM. However, while researching the semantic …