GraphNews

ServiceNow is joining Open Semantic Interchange (OSI)! Sharing a few of my personal thoughts: 🟢 Three years ago at Snowflake Summit 2023, I was making the case that LLMs couldn’t just “figure out” the semantics of your business by giving the model training data, fine-tuning it, etc. And that is why we need to invest in semantics and knowledge graphs. The pushback was: show us the numbers. So we built the benchmark. Our research result was one of the first to rigorously demonstrate the accuracy gap between LLMs querying raw data versus LLMs equipped with an ontology and knowledge graph. That accuracy increase was 3X at that time. It helped wake up the industry to what knowledge graphs and ontologies actually bring to the table. 🟢 Across hundreds of recent customer conversations, I’m watching organizations realize that the semantics of your business, your definitions, your metrics, your processes, your decisions, are the DNA of your company. Frontier models don’t come with that. You have to give it to them. And if it’s not standardized from the start, you’re going to rebuild it over and over again. The vendor lock-in concern is real. 🟢 There are actually two different problems here, and people are conflating them. 1. Semantic interoperability as infrastructure. This is the work of defining your business concepts, metrics, processes, decision logic in a way that is open, interoperable, and machine-readable. It’s foundational. It’s about ensuring that when any system (AI or otherwise) touches your business data, operational or analytical, and it operates from the same shared understanding. 2. Semantic context for AI agents. AI agents need to understand the context of the business to act on your behalf. This is a powerful, urgent use case, and it’s creating the most powerful organizational incentive we’ve ever had to finally invest in semantics. Don’t confuse the consumer (AI agents) with the foundation (semantic interoperability). Both matter. My suggestion is to start small and think big with “Semantic interoperability as infrastructure” as the goal because “Semantic context for AI agents” will come with it. If you focus on the goal of AI agents first, then you are missing out and most probably end up building brittle systems. 🟢 If you’re just getting started: don’t try to model your enterprise on day one. Start with your data warehouse semantics, your metrics definitions, your BI layer. Build governance around it. Then expand. 🟢 Proud that the og data.world team’s work contributed to putting semantics in the spotlight, and now honored to now be contributing to OSI alongside so many great partners. 🟢 For those asking about how OSI relates to the broader semantic web and W3C standards (i.e. RDF/OWL/SHACL),  yes, that thread is coming. Is your investment in semantics for AI agent requirements, or are you thinking about it as broader semantic interoperability as infrastructure? Link to my substack and ServiceNow announcement blog | 38 comments on LinkedIn

Part 1. Key Definitions Vocabulary A controlled set of terms (labels) used to describe a domain.

Having your master data in RDF is often not an option and this is where a hazardous search for ETL/sync/materialization starts. I wrote before about YARRRML, maplib, RML, Morph-KG, LinkML and similar tools. Here I'd like to highlight the virtual graph approach: pulling SQL via SPARQL and turning Postgres into a triple store via extensions. pgRDF is a Rust-native Postgres extension built on pgrx that stores RDF as a dictionary-encoded hexastore. The benchmarks are unusually ambitious, SHACL validator and an OWL 2 RL reasoner. Quite impressive. The catch is the single dev and maturity of the project. Ontop is a popular and reliable option and has been integrated into GraphDB as a data virtualization solution. It's solid. Caveats live in some specific needs like heavy usage of OPTIONAL, aggregation heavy queries and highly dynamic data. rdf_fdw is a Postgres foreign data wrapper that enables seamless integration with RDF triplestores via SPARQL endpoints. It supports querying RDF data using SQL, somewhat like ontop. Here as well, I'd be careful with using a Github project that has not proven itself. The more enterprisey option involves messaging (Kafka) and workflows (Kestra, Prefect...) with the added advantage that it gives you a company-wide integration rather than a one-to-one pipeline. Plus, this works with other relational DBs, not just Postgres. Did I overlook a major/magical option? - pgRDF: https://lnkd.in/epXw6g2R - ontop: https://ontop-vkg.org - rdf_fdw: https://lnkd.in/ei8etVCQ #RDF #GraphDatabase #KnowledgeGraphs

For those of us still wondering about the Kuzu post-mortem: ` I Analyzed 163k Lines of Kuzu’s Codebase. Here’s Why Apple Wanted It. ` tl;dr: Apple needed a graph database to make Apple Intelligence, you know...intelligent and on device. It's the same story we're seeing again and again: Throwing LLMs at something by themselves will result in slop, hallucinations, and silver-tongued vapors. Graph technology provides a backbone to knowledge, intelligence, and in Kuzu's case, a bit of wisdom. This article is obviously just the author's personal perspective. I'm happy to hear your take too. https://lnkd.in/gDvykhdS #graphdatabases #kuzu

It seems like we're fast approaching a context hangover. The first half of the year, you couldnt attend a conference, login to LinkedIn, or browse a vendor website without being constantly bludgeoned with references to context. This is still the case, but the hype seems to be fading. Why? Because the idea of 'context' is too academic for most CDO's. Our discussions are too focused on the technicalities of things like ontologies and semantic layers. There's almost no focus on *why* these technologies are important, or what the business value of implementing them will be. The value of context is consistently expressed through capabilities, and not business outcomes. This only magnifies the confusion, and I suspect most are thoroughly exhausted by it. On the bright side, we're on the downslope of all the hype around context. The hangover is next. #context #semanticlayer #knowledgegraph | 47 comments on LinkedIn

An ontology needs a way to evolve without pretending its past never existed. I’ve been working through this problem while designing lifecycle governance for semantic entities. The central distinction I keep returning to is: Unused is a signal. Obsolete is a governance decision. Usage data can help identify candidates for retirement: classes with no instances, properties nobody queries, or entities that no longer receive new relationships. But before deleting anything, we still need to ask: - Does historical data depend on it? - What relationships or downstream systems reference it? - Has another entity replaced it? - Would removing it erase part of the system’s semantic history? In many cases, the safer outcome is governed deprecation: - Remove the entity from the active semantic model - Prevent new relationships or uses - Preserve its historical relationships, provenance, and lineage - Identify what replaced it - Record why, when, and by whom it was deprecated This changes the architectural question. It is no longer simply, “Can this entity be deleted?” It becomes, “How can we simplify the active model without breaking dependencies or erasing the knowledge needed to interpret its history?” #SemanticEngineering #DataGovernance #AppliedOntology #KnowledgeGraphs #OntologyStudio

Vector RAG, GraphRAG, and Agentic RAG preserve different context. Learn when each works, where each fails, and how to combine them in one app.

Storing your context in a graph can be the right call. Hooking that graph up to the agent almost never is, and the two decisions get made as though they were one.

The ecosystem around ontologies, OWL, and semantic reasoning has experienced a major revival. This resurgence is largely driven by Neuro-Symbolic AI, RAG (Retrieval-Augmented Generation), and the need for deterministic knowledge grounding to eliminate LLM hallucinations.

Why one ontology no longer fits every workload

Landmark exit for Integral Integral Capital Group has agreed to sell Graphwise, a pioneer in knowledge graph and semantic AI technology, to Oakley Capital - marking one of the largest software exits in Bulgaria's history. Oakley's Fund VI will acquire a majority stake, partnering with Graphwise's founders and management to support the Company's continued growth. The exit caps a journey that began in 2022, when an Integral-led consortium - alongside Portfolion Capital Partners, Carpathian Partners LLP and the EBRD, with European Investment Fund (EIF) as co-investor - acquired Ontotext, the Sofia-founded knowledge graph pioneer founded by Atanas Kiryakov. In 2024, Ontotext merged with Vienna's Semantic Web Company to form Graphwise, now serving more than 200 blue-chip customers worldwide. Graphwise's technology sits at the heart of one of enterprise AI's central challenges: grounding large language models in accurate, governed data. As organisations move from AI experimentation to production, that positioning has never been more strategically relevant. This transaction underscores Integral's ability to build category-leading technology businesses in the CEE region - from platform creation through to a landmark exit. Thank you to everyone we built Graphwise together with and who brought this exit home: - Graphwise founder Atanas Kiryakov - The management team: Andreas Blumauer, Martin Kaltenböck, Vassil Momtchev, Veska Sergieva-Davidova, Rick Crane, Todd Blaschka and Martin Kelchev - Our co-investors at PortfoLion Capital Partners and Carpathian Partners - The entire Graphwise team Arma Partners advised the selling shareholders and Graphwise, with Deloitte, Boston Consulting Group (BCG), DGKV and Gowling WLG providing due diligence and legal support. #PrivateEquity #EnterpriseAI #KnowledgeGraph #Bulgaria #CEE #MandA Evtim Chesnovski Ivan Abadjiev, CFA | 19 comments on LinkedIn

I am extremely pleased to announce that we have officially brought Oakley Capital on board as a new investor and majority shareholder in Graphwise! Oakley is a leading European private-equity investor focused on mid-market companies and was founded 20 years ago by Peter Dubens to serve as the partner of choice for outstanding founders and entrepreneurs. This acquisition will accelerate Graphwise's global expansion and solidify knowledge graphs as critical infrastructure in the broader AI market, enabling enterprises to deploy highly accurate, trusted, and cost-effective AI systems. AI is fundamentally changing the way every company operates, making the need for trustworthy, well-structured data more important than ever. The true value of AI depends on the quality and reliability of the underlying data. That’s why we developed Graphwise—to provide companies with their “semantic backbone,” enabling large language models (LLMs) and agents to build upon verified business data, thereby overcoming the challenges of factual consistency and rapidly rising costs that so many companies face today. Our partnership with the outstanding team at Oakley Capital marks a crucial milestone. They understand our technology inside and out and share our goal of helping companies scale AI cost-effectively. Thanks to their exceptional track record in transforming founder-led software companies, we are now ideally positioned to accelerate our growth, expand our presence in key international markets, and thereby strengthen our market-leading position. A huge thank you to our leadership team—including our President, Atanas Kiryakov—our dedicated Graphwise team, our partners, our investors & advisors, and the more than 200 prominent customers who have placed their trust in us along this journey. Here’s to the next exciting chapter in building the most trusted infrastructure for enterprise AI! 🥂 #Graphwise #OakleyCapital #EnterpriseAI #KnowledgeGraphs #SemanticBackbone #SemanticLayer #GraphRAG #AI #TechNews #Growth

If an agent can learn an ontology from its own memory, and that ontology then shapes what it learns next, where does memory end and a world model begin? The usual story is that ontologies ground agents. But once agents are able to learn and enrich those ontologies from their own traces, you have a f

"We need an ontology." "Isn't that just a knowledge graph?" Five terms, used almost interchangeably, that actually describe distinct layers of a knowledge architecture, from "how do we name and organise things" up to "what does an AI agent actually load into its context window.

Quick introduction to context graphs We’ve all heard about context graphs: concepts linked by clearly defined relationships that represent some domain (like retail or agriculture). Context graphs’ purpose is to provide context to users—human or machine agents—so that they can take grounded, meaningf

How to increase ROI, or create new revenue streams, from your graph data without selling the whole thing? That's a question many of us have been struggling w...

We describe Research Assistant, an internal LLM-based system developed at AstraZeneca to help scientists and clinicians explore biomedical questions across a broad range of data sources. The...

Sign in or join now to see posts like this one and more.