<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Ontology + Knowledge Graph on Medium]]></title>
        <description><![CDATA[Stories by Ontology + Knowledge Graph on Medium]]></description>
        <link>https://medium.com/@steventseng.ai?source=rss-261ec68d8144------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*g4--7-_zIDiFc0-A1CX0iQ.jpeg</url>
            <title>Stories by Ontology + Knowledge Graph on Medium</title>
            <link>https://medium.com/@steventseng.ai?source=rss-261ec68d8144------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 22:18:42 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@steventseng.ai/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Why Ontology Is Becoming the Core Layer of the Enterprise AI Stack]]></title>
            <link>https://medium.com/@steventseng.ai/why-ontology-is-becoming-the-core-layer-of-the-enterprise-ai-stack-bb1a6602eb4b?source=rss-261ec68d8144------2</link>
            <guid isPermaLink="false">https://medium.com/p/bb1a6602eb4b</guid>
            <category><![CDATA[ontology]]></category>
            <category><![CDATA[knowledge-graph]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[enterprise-ai]]></category>
            <category><![CDATA[graphrag]]></category>
            <dc:creator><![CDATA[Ontology + Knowledge Graph]]></dc:creator>
            <pubDate>Fri, 20 Nov 2026 04:44:49 GMT</pubDate>
            <atom:updated>2026-11-20T04:45:46.442Z</atom:updated>
            <content:encoded><![CDATA[<blockquote>An insurance company asked a RAG system about “policy” and received three plausible documents. The problem was not the model. It was the representation.</blockquote><p>An insurance company put its documents into a RAG system and asked about “policy.”</p><p>The system returned three chunks: a customer insurance policy, an underwriting policy, and a compliance policy. All three scored well on textual similarity. Only one answered the question.</p><p><a href="https://enterprise-knowledge.com/ontology-and-knowledge-graph-in-the-age-of-ai-and-agents/">Enterprise Knowledge used this example in a June 2026 article to make a point that keeps resurfacing in the knowledge graph community.</a> An ontology represents each type of policy as a distinct concept, with distinct relationships. That context is what lets retrieval tell them apart.</p><p>It looks like a small example. It points to a larger shift.</p><h3>Ontology is not data</h3><p>The distinction matters.</p><p>Enterprise Knowledge describes an ontology as a semantic schema: a formal model of the entities in a domain, their attributes, the relationships between them, and the constraints on those relationships.</p><p>A knowledge graph is data structured through that ontology. The ontology says what a <em>Customer</em> and a <em>Purchase Order</em> mean. The graph holds a particular customer, a particular order, and the link between them.</p><p><a href="https://atlan.com/know/ai-agent/knowledge-graph/ontology-vs-knowledge-graph/">Atlan puts it more bluntly: enterprise AI needs both.</a> The ontology tells agents what concepts mean. The knowledge graph tells them what exists. Remove either side and an agent has either no shared vocabulary or nothing concrete to query.</p><p>This ordering matters when teams decide what to build first.</p><p>Putting documents into a vector store is fast. But a document does not tell the system what to do when someone cancels an order. That is a question about meaning, not retrieval.</p><h3>Three signals</h3><p><strong>The vocabulary has changed.</strong> <a href="https://www.youtube.com/watch?v=71i-tjPz7bI">A session titled <em>2026 Knowledge Graph, AI, and Other Data Trends</em> includes a chapter called “2026, The Year of the Ontology.”</a> The same session opens with what it calls the ontologist job market paradox: demand is rising, yet companies must grow the people who can do the work internally.</p><p><strong>Enterprise forecasts followed.</strong> <a href="https://www.linkedin.com/posts/tianyouxu_ontology-ai-realestate-share-7400549585074671616-aJDd/">A LinkedIn post framed its 2026 prediction in one sentence: ontology will become a core focus of the enterprise AI stack</a>. <a href="https://atlan.com/know/ai-agent/knowledge-graph/knowledge-graph-construction-for-ai/">Atlan released a guide on <em>Knowledge Graph Construction for AI</em></a>. In November, <a href="https://graphwise.ai/event/building-a-semantic-backbone-for-agentic-ai-over-enterprise-data/">Graphwise demoed <em>Building a Semantic Backbone for Agentic AI over Enterprise Data</em></a>.</p><p>These events have moved past explaining what a knowledge graph is. They are now about making one hold up in production.</p><p><strong>The conference floor agrees.</strong> After the seventh Knowledge Graph Conference at Cornell Tech, <a href="https://medium.com/@giuseppefutia/notes-from-kgc-2026-c9b4ac8569e5">attendee Giuseppe Futia wrote down two lines that kept coming up</a>:</p><ul><li>Knowledge graphs are the substrate on which production AI stands.</li><li>The real leverage is representation, not prompting.</li></ul><p>His summary of the week was direct. Most failures blamed on model limitations or prompt engineering are actually representation failures.</p><h3>What this changes in practice</h3><p>If ontology becomes a core layer, the order of work changes with it.</p><p>Graph RAG is the most visible application. Enterprise Knowledge notes that it retrieves over a structure governed by an ontology, rather than over flat text chunks. That is what resolves the “policy” ambiguity above.</p><p>There is also a cost argument. Every irrelevant chunk pulled into context costs tokens. Classifying data through an ontology lets an agent retrieve only the type it needs.</p><p>Two <a href="https://www.linkedin.com/posts/kgc-2026-kicks-off-with-two-days-of-hands-on-share-7450559726246952964-CE5P/">workshops at KGC 2026</a> described a second shift more clearly.</p><p><a href="https://watch.knowledgegraph.tech/videos/ai-assisted-ontology-engineering-guardrails-and-workflows-for-using-llms-safely">One covered AI-assisted ontology engineering</a>. The diagnosis was direct: used naively as ontology generators, LLMs produce tangled class hierarchies, inconsistent taxonomies, and runaway property growth. A larger model does not fix this. The problem is governance.</p><p>The proposed solution reverses the relationship. The ontology stops being an LLM output. It becomes the harness around the LLM: a declarative constraint on what the agent may reason about, propose, or change. A seven-stage review workflow supports it, along with immutable audit trails and quality gates.</p><p><a href="https://watch.knowledgegraph.tech/kgc-2026/videos/bridging-rdf-and-property-graphs">Another workshop tackled translation between RDF and property graphs</a>. The usual pattern is a one-way migration that quietly drops meaning. The proposed fix is a <em>fidelity contract</em>: an explicit, testable statement of what must be preserved, what may be projected, and how to prove it.</p><p>Graph databases and RDF stop being an either-or choice, as long as movement between them has terms.</p><h3>What it does not solve</h3><p>Ontologies are expensive to maintain. Defining concept boundaries requires someone who understands both the domain and the model. The “ontologist job market paradox” is another way of saying that these people are not sitting on a shelf.</p><p>Speed is not the measure of maturity. The signals <a href="https://medium.com/@giuseppefutia/notes-from-kgc-2026-c9b4ac8569e5">Futia emphasises from KGC</a> are unglamorous: audit trails, validation gates, provenance, and conformance scoring. Those are what separate a demo from a system that can operate in a public-facing environment.</p><p>Not every problem needs this layer. If a model already answers reliably and time does not change the answer, there is no reason to build a semantic scaffold first.</p><p>The cases that justify it share three features: costly errors, concepts that are easy to confuse, and a need to explain how a conclusion was reached.</p><h3>Where to start</h3><p>For a team adding a data layer beneath an agent, begin with one narrow domain.</p><ul><li>Write down three to five core concepts and their relationships.</li><li>Give RAG a skeleton to check against.</li><li>Add rules and validation incrementally.</li></ul><p>Treat the ontology as a contract you expect to revise. Do not treat it as a design you must finish before anything ships.</p><p>Ontology is getting attention this year for a reason. Agents are finally running into problems that the knowledge graph community has handled for two decades.</p><p>Concepts need meaning. Relationships need to be queryable. Decisions need to be traceable.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lr2CPudM56TlcrxsdsNWOA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*x1-mPJ6gloTCY-ibNowIPw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_2i1VvSPmMMb5K3PPUqjYg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XJNgvqzXWQ0J0QlDZNAw-Q.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bb1a6602eb4b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Scar Tissue Behind Every Knowledge Graph]]></title>
            <link>https://medium.com/@steventseng.ai/the-scar-tissue-behind-every-knowledge-graph-a16d1767fac6?source=rss-261ec68d8144------2</link>
            <guid isPermaLink="false">https://medium.com/p/a16d1767fac6</guid>
            <category><![CDATA[knowledge-graph]]></category>
            <category><![CDATA[enterprise-ai]]></category>
            <category><![CDATA[semanticweb]]></category>
            <category><![CDATA[data-governance]]></category>
            <category><![CDATA[ontology]]></category>
            <dc:creator><![CDATA[Ontology + Knowledge Graph]]></dc:creator>
            <pubDate>Wed, 18 Nov 2026 01:34:48 GMT</pubDate>
            <atom:updated>2026-11-18T01:34:48.994Z</atom:updated>
            <content:encoded><![CDATA[<p>Twenty years, twenty lessons, and why most of the work has nothing to do with graphs.</p><p>This week at <a href="https://www.kmworld.com/Articles/News/News/Lessons-learned-from-two-decades-of-building-ontologies-and-knowledge-graphs-at-KMWorld-2026-177540.aspx">KMWorld 2026 in Washington, D.C., Juan Sequeda gave a talk titled <em>Scar Tissue: 20 Lessons from 20 Years of Building Ontologies and Knowledge Graphs</em>.</a> The title is the point.</p><p>Knowledge graphs and ontologies are suddenly everywhere in AI conversations. The people who have worked in this field since 2005 do not remember a smooth ascent. They remember scars.</p><p>Sequeda is now a principal researcher at ServiceNow and a co-founder of data.world. Over two decades, he has moved this work from academic papers into products and enterprises. His starting claim is simple: knowledge is becoming a first-class citizen. Knowledge graphs can help organisations scale understanding, adapt to change, and make AI useful for the business.</p><p>That is the promise. The twenty lessons describe what it takes to reach it.</p><p>Sequeda organises the lessons as four stages of maturity, rather than a checklist. Each stage has its own delusion. The sequence matters because most teams are somewhere in stage one and do not know it.</p><h3>Stage 1: “We thought this was mostly a technical problem.”</h3><p>The first stage is naïve optimism. Pick a graph database, model some entities, load the data, and expect the system to work.</p><p>Sequeda’s first lessons explain why this fails.</p><p><strong>Knowledge graphs make LLMs more accurate.</strong> That is why the topic has become suddenly popular. But you cannot automate understanding. You can automate extraction, mapping, and validation. Understanding is the part where someone takes responsibility for what a concept means.</p><p>An LLM can produce a taxonomy that looks correct and means nothing.</p><p>A related trap is assuming that a knowledge graph built from structured data resembles one built from text. Structured data gives you explicit semantics. Text requires you to recover them. Teams often treat the two as interchangeable because both eventually appear as nodes and edges on a screen.</p><p>The advice in this stage is practical.</p><p><strong>Ontology is a means, not a goal.</strong> <strong>Do not be pedantic.</strong> <strong>Do not try to do everything.</strong></p><p>The urge to design a complete, internally consistent model of the business before loading a single row is the most reliable way to ship nothing. A smaller model that answers a real question beats a perfect model that lives on someone’s laptop.</p><h3>Stage 2: “Oh, this is harder than we thought.”</h3><p>Reality arrives when the first model meets the second department.</p><p>The schema that made sense to one team collides with another team’s vocabulary. The project discovers that its real problem is not graph traversal.</p><p>You need a methodology: a repeatable way to move from question to model, from model to data, and from data to validation. This is socio-technical work. The hardest parts involve people who use different words for the same thing, and the same word for different things.</p><p>Two lessons matter here.</p><p>First, use cases for knowledge graphs should focus on <strong>reusability and managing uncertainty</strong>, rather than producing one-off answers. A graph that answers one question is useful. A graph that helps the organisation answer a family of related questions is an asset.</p><p>Second, teams must choose between <strong>application-centric and data-centric thinking</strong>. Should the graph serve a specific decision, or should it model the entire enterprise up front? Sequeda’s advice is to start with the application. Let reuse pull the model outward.</p><p>Then comes the lesson teams almost always skip: <strong>define the incentive mechanism</strong>.</p><p>Why would the person who owns the source data maintain the mapping? Why would the domain expert keep the taxonomy current? A knowledge graph that depends on unpaid goodwill has an expiry date.</p><p>Governance debt accumulates quietly, then all at once.</p><h3>Stage 3: “The problem is not data. It is meaning.”</h3><p>Semantic awakening is where the work stops being about pipelines and starts being about authorship.</p><p><strong>You need knowledge engineers.</strong> These people sit between a domain expert and a formal model. They understand the business and the semantics. This is not a role you can create by renaming a data engineer and expecting the same output.</p><p><strong>Governance is what gets you out of the proof of concept.</strong> A proof of concept proves that the technology works. Governance, including ownership, review, change control, and accountability, makes the system safe to rely on.</p><p>This explains why so many impressive demos never reach production. The demo did not fail. The organisation around it was never built.</p><p>Two technical lessons carry most of the weight in this stage.</p><p><strong>Identity is the root of integration problems.</strong> You must be able to say that the “account” in CRM and the “account” in billing are the same thing. You must also prove it, stably, over time. If you cannot, every downstream join is a guess.</p><p><strong>Mappings are where semantics live.</strong> The meaning of an ontology does not sit in its class names. It sits in the mappings that connect those classes to the data and to each other.</p><p>This is the argument against buying a knowledge graph off the shelf and expecting AI agents to understand it. The hard part is not storing relationships. The hard part is deciding what they mean, then maintaining that decision as the business changes.</p><h3>Stage 4: Knowledge infrastructure</h3><p>The final stage is where the work becomes sustainable.</p><p><strong>Your first knowledge graph should be about your metadata.</strong> Before you model customers and products, model what you already have: datasets, columns, owners, lineage, and definitions. This is the lowest-risk place to learn the discipline, and it can pay for itself quickly.</p><p><strong>Separate knowledge from data.</strong> The model of what a thing means should not be tangled up with the rows that instantiate it. Mixing the two is how a graph becomes a very expensive reporting layer.</p><p><strong>A knowledge graph is not a graph database.</strong> The database is storage. The graph is a claim about the world. You can store a graph in a relational database and still have a knowledge graph. You can also store millions of triples and have no knowledge graph at all.</p><p>Then there is the lesson that should be printed on the wall of every semantic technology conference:</p><p><strong>RDF versus property graph is a distraction. Stop.</strong></p><p>Sequeda has argued on both sides of that debate. He has also helped build bridges between the two communities. That work includes PG-Schema, which won a SIGMOD best paper award and contributed to ISO GQL.</p><p>His position is clear. Property graphs are excellent databases. RDF is excellent for knowledge representation. The two are not mutually exclusive, and convergence is already underway.</p><p>The debate confuses newcomers, stalls adoption, and moves the industry backwards. Treat knowledge as first-class. Everything else is secondary.</p><p>The final lesson is the most honest: <strong>we still lack tooling</strong>. After twenty years, the discipline still depends on a patchwork of editors, validators, and scripts. Anyone who tells you the tooling problem is solved is selling something.</p><h3>What this means for AI agents</h3><p>This is why an old talk is worth reading in November 2026.</p><p>Enterprises are buying knowledge graphs for their AI agents, then skipping the layer that makes those graphs work. They want reasoning without semantics, retrieval without mappings, and governance without meetings.</p><p>Sequeda’s account says that order cannot be skipped.</p><p>Understanding cannot be automated. Ontology is a means. Governance is what gets an organisation out of the proof-of-concept stage. Identity is where integration hurts. Mappings are where meaning lives.</p><p>These are not caveats on the way to a knowledge graph. They are the work.</p><h3>A practical test</h3><p>Before funding another knowledge-graph pilot, ask five questions.</p><ul><li>Who owns each core concept, and who can change its definition?</li><li>Which team maintains the mappings between source systems and the ontology?</li><li>How does the system prove that two records refer to the same real-world entity?</li><li>What happens when a source system changes its schema, owner, or business rules?</li><li>Which decision, workflow, or agent becomes better because this graph exists?</li></ul><p>If the answers depend on goodwill, undocumented conventions, or a single person’s institutional memory, the project is still at stage one.</p><p>The good news is that the destination is real. Knowledge graphs can help organisations scale understanding, adapt to change, and make AI useful for the business. Twenty years of scar tissue is a reasonable price for knowing how to get there without repeating every mistake.</p><p>The bad news is that many teams are still at stage one, calling it production.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ngIg1SMgw2D3LVy8XHJy-A.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7owOovJ0U-L6V4ANPmtPRw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pbqMDMefH4kcGAGGmD7Xlw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PoVaU0o625PrUhq-lvUm5A.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Nh5Xy88uDLG8XPYuV3a2xw.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a16d1767fac6" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>