Understanding Graph Technologies

In the AI era, your database isn’t just a backend choice — it’s a strategic enabler. AI systems today are not just consuming data. They're reasoning over it, retrieving it, embedding it, and traversing relationships across it. And that changes everything about how we choose databases. Here’s a side-by-side comparison I created to show how different databases align with modern AI workloads: • 𝗥𝗲𝗹𝗮𝘁𝗶𝗼𝗻𝗮𝗹 𝗗𝗕𝘀 — Still critical for structured systems (ERP, Finance), but struggle with unstructured and high-dimensional data.    • 𝗡𝗼𝗦𝗤𝗟 𝗗𝗕𝘀 — Great for flexible, high-throughput ingestion (IoT, real-time analytics), but limited for complex joins and semantic context.    • 𝗩𝗲𝗰𝘁𝗼𝗿 𝗗𝗕𝘀 — The core of GenAI. They make semantic search, embeddings, and RAG architectures possible.    • 𝗚𝗿𝗮𝗽𝗵 𝗗𝗕𝘀 — Ideal for modeling relationships, reasoning, and powering agent memory and decision graphs. In the AI-native stack, Vector and Graph databases are foundational: • LLMs retrieve semantically matched chunks via vector search    • Agents reason through graph traversals and decision paths    • Hybrid models use all four — ingesting via NoSQL, storing core logic in relational, retrieving via vector, and reasoning via graph.   It’s not just about data storage — it’s about enabling intelligence.

I recently worked on a project that explores how structured data, graph systems, and evaluation frameworks can come together to build reliable AI-driven applications in the nutrition domain. I curated and processed the USDA FoodData Central dataset to extract over 260,000 semantic Knowledge Graph triplets, capturing relationships between food items, nutrients, measurement units, and categories. These triplets were then used to construct a Knowledge Graph using Neo4j. To make the graph interactable and verifiable, I integrated it with the Graphiti framework to build a retrieval-based agent capable of answering health and nutrition-related questions grounded in the underlying data. To assess the reliability of the agent's responses, I used the `judgeval` library from Judgment Labs. The evaluation involved: 1) Instruction adherence unit tests 2) Faithfulness scoring to detect hallucinated or unsupported answers 3) Benchmarking against the NutriBench dataset, which provides realistic meal descriptions and expected nutrient-level outputs. The full pipeline—ranging from dataset curation to KG construction, agent design, and benchmark evaluation—is documented here: https://lnkd.in/gxNg7CjS I would be glad to hear thoughts or feedback from others working at the intersection of knowledge graphs, ontologies, agentic evaluation, and applied AI in healthcare and nutrition. #KnowledgeGraph #Neo4j #Graphiti #Judgeval #LangGraph #LLMAgents #AIValidation #NutritionInformatics #NLP #OpenSource #NutriBench

A hydraulic system fails on a production line. The maintenance team spends eight days running diagnostics. New parts, new tests, more downtime. Bob van de Kuilen 's team loaded the PLC code and P&IDs into a knowledge graph, asked Claude to trace the relationships, and found the root cause in three minutes. A contaminated proportional valve. Eight days vs. three minutes. Same data. Same equipment. The difference was context. Bob joined us on Unplugged: An IIoT Podcast to explain why knowledge graphs pick up where UNS leaves off. UNS organizes your data into a tree. Knowledge graphs connect it into a web of meaning, the way your brain actually works when solving problems. One sensor on a production line means three different things to three different teams. Supply chain counts WIP. Reliability tracks running hours. Production feeds OEE. A knowledge graph holds all three perspectives on the same data point. UNS cannot do that. If you are building a data strategy and skipping human context, you are building a library with no librarian. Full episode in comments. #IIoT #KnowledgeGraphs #Manufacturing #IndustrialData #DigitalTransformation

I’ve had 250 hours of calls with engineering leaders over the last 12 months. The #1 thing I learned? Nearly every team has tried to build an AI agent to help manage their cloud infra, and nearly every team has failed. Here’s why. 👇 Infrastructure is too complex. Teams are too small. And incidents don’t wait. You’re juggling 15+ microservices, a mountain of metrics, a half-written runbook, and 3 tabs of logs. All while trying to reduce MTTR without burning out your best engineers. Everyone wants an AI agent to help. But they all hit the same wall: * Information about your infrastructure and its context is often buried in docs, Slack threads, or in engineers’ heads where AI agents would have no hope of ever accessing it. * AI agents don’t work well without infrastructure context. * That context can include highly sensitive data. * No team wants to ship that sensitive data to a black-box vendor. We’ve tried building a root cause analysis agent ourselves. And it worked. But only after we build a real-time infrastructure map (we call it a Knowledge Graph) of all our systems, complete with dependencies, resource metadata, and metrics APIs. So then we realized… 🤔 This knowledge graph is useful for way more than just RCA ✋ Other teams want to use it too, but they need to run it locally 🤷 No existing tools (Backstage, Port, etc.) have been built for AI-first automation So that’s why we’re open sourcing the knowledge graph that powers our AI SRE Agent. * It builds a real-time graph of your infra from AWS environments and other sources. * It runs locally and keeps your data private. * It exposes an MCP server so you can plug it into Cursor, Claude, Copilot, whatever AI tools you’re already using. We think this is the missing piece for anyone serious about building AI agents for infrastructure management that are actually useful. It’s not just a chat interface, and it’s not just producing a static software catalog that you have to maintain. It’s a programmable, agentic foundation for automation, built on YOUR real system context. If you want early access to the GitHub repo, you can sign up using the link in the comments 📌

In the AI arms race, data isn’t just fuel - it’s the architecture for the intelligence you train. Yet most enterprises still rely on 20th-century data architectures for 21st-century intelligence. Your CRM is a vault of customer interactions, your ERP tracks orders, and your analytics tool crunches numbers - each a walled garden. AI is meant to be the brain that connects them all, but it can’t - because these systems weren’t designed for AI. Relational databases, JSON APIs, and vector embeddings only scratch the surface. Structured isn’t the same as meaningful, and fragmented data doesn’t tell the full story. AI doesn’t just need data - it needs context, connections, and meaning. 🔵 Why Knowledge Graphs? Knowledge Graphs (KGs) do not really store information - they allow AI to understand it. Instead of scattered, messy data, KGs create an interconnected web of meaning, giving AI the depth it needs to make informed, explainable decisions. Think of a KG as a mind map for your entire organisation. A customer isn’t just a database row; they’re linked to past purchases, support tickets, email exchanges, written notes, social sentiment, and pricing preferences. An insurance claim isn’t just an entry - it’s tied to policy details, vehicle history, repair records, and similar cases. This isn’t about storage - it’s about making sense of complexity at a scale that rigid databases and APIs simply can’t match. 🔵AI Needs Meaning, Not Just Data: Large Language Models generate plausible-sounding responses - but they lack deep domain expertise. That’s where ontologies come in - structured vocabularies that teach AI what concepts actually mean in a specific context. Want an AI that actually gets your financial reports? Or a recommendation engine that pinpoints the perfect match for your products? The trick isn’t just data - it’s meaningful structure. Ontologies aren’t mere schemas; they’re models of meaning, mapping out your domain in formal logic. That precision makes your data verifiable, and verifiable knowledge is where AI thrives. 🔵 AI Needs Connected Data: To take advantage of scaling laws you need to bring all your data together, but your data isn’t in one place anymore. It’s scattered across spreadsheets, CRMs, APIs, legacy databases, documents, and emails. AI needs to see the whole picture to act intelligently. KGs act as a semantic layer that links these sources into a single source of truth - without moving the data. The web itself is a graph. LLMs were trained on the web. If you want AI that understands your organisation, your data needs to be connected the same way. 🔵 The Bottom Line: Knowledge Graphs aren’t just a new way to structure data - they’re a new way to think about your organisation’s knowledge. If your AI initiatives lack direction, structure, reliability, explainability, flexibility, or interoperability, the problem isn’t AI - it’s how you’re thinking about your data. ⭕ KGG: https://lnkd.in/ezHU2amU

One year ago today, Dean Allemang Bryon Jacob and I released our paper "A Benchmark to Understand the Role of Knowledge Graphs on Large Language Model's Accuracy for Question Answering on Enterprise SQL Databases" and WOW! Early 2023, everyone was experimenting with LLMs to do text to sql. Examples were "cute" questions on "cute" data. Our work provided the first piece of evidence (to the best of our knowledge) that investing in Knowledge Graph provides higher accuracy for LLM-powered question-answering systems on SQL databases. The result was that by using a knowledge graph representations of SQL databases achieves 3X the accuracy for question-answering tasks compared to using LLMs directly on SQL databases. The release of our work sparked industry-wide follow-up: - The folks at dbt, led by Jason Ganz, replicated our findings, generating excitement across the semantic layer space - Semantic layer companies began citing our research, using it to advocate for the role of semantics - We continuously get folks thanking us for the work because they have been using it as supporting evidence for why their organizations should invest in knowledge graphs - RAG got extended with knowledge graphs: GraphRAG - This research has also driven internal innovation at data.world forming the foundation of our AI Context Engine where you can build AI apps to chat with data and metadata. Over the past year, I've observed two trends: 1) Semantics is moving from "nice-to-have" towards foundational: Organizations are realizing that semantics are fundamental for effective enterprise AI. Major cloud data vendors are incorporating these principles, broadening the adoption of semantics. While approaches vary (not always strictly using ontologies and knowledge graphs), the message is clear: semantics provides your unique business context that LLMs don't necessarily have. Heck, Ontology isn't a frowned upon word anymore 😀   2) Knowledge Graphs as the ‘Enterprise Brain’: Our work pushed to combine Knowledge Graphs with RAG, GraphRAG, in order to have semantically structured data that represents the enterprise brain of your organization. Incredibly honored to see Neo4j Graph RAG Manifesto citing our research as critical evidence for why knowledge graphs drive improved LLM accuracy. It's really exciting that the one year anniversary of our work is while Dean and I are at the International Semantic Web Conference. We are sharing our work on how ontologies come to the rescue to further increase the accuracy to 4x (we released that paper in May). This image is an overview of how it's achieved. It's pretty simple, and that is a good thing! I've dedicated my entire career (close to 2 decades) to figure out how to manage data and knowledge at scale and this GenAI boom has been the catalyst we needed in order to incentivize organizations to invest in foundations in order to truly speed up an innovate. There are so many people to thank! Here’s to more innovation and impact!

𝐄𝐯𝐞𝐫𝐲𝐨𝐧𝐞 𝐭𝐚𝐥𝐤𝐬 𝐚𝐛𝐨𝐮𝐭 𝐑𝐀𝐆 (𝐑𝐞𝐭𝐫𝐢𝐞𝐯𝐚𝐥-𝐀𝐮𝐠𝐦𝐞𝐧𝐭𝐞𝐝 𝐆𝐞𝐧𝐞𝐫𝐚𝐭𝐢𝐨𝐧) 𝐥𝐢𝐤𝐞 𝐢𝐭 𝐢𝐬 𝐣𝐮𝐬𝐭 𝐨𝐧𝐞 𝐭𝐡𝐢𝐧𝐠. 𝐁𝐮𝐭 𝐡𝐞𝐫𝐞 𝐢𝐬 𝐭𝐡𝐞 𝐫𝐞𝐚𝐥𝐢𝐭𝐲: 𝐭𝐡𝐞𝐫𝐞 𝐚𝐫𝐞 𝐝𝐢𝐟𝐟𝐞𝐫𝐞𝐧𝐭 𝐟𝐥𝐚𝐯𝐨𝐫𝐬 𝐨𝐟 𝐑𝐀𝐆 𝐞𝐚𝐜𝐡 𝐬𝐨𝐥𝐯𝐢𝐧𝐠 𝐝𝐢𝐟𝐟𝐞𝐫𝐞𝐧𝐭 𝐩𝐫𝐨𝐛𝐥𝐞𝐦𝐬 𝐢𝐧 𝐡𝐨𝐰 𝐀𝐈 𝐫𝐞𝐭𝐫𝐢𝐞𝐯𝐞𝐬, 𝐨𝐫𝐠𝐚𝐧𝐢𝐳𝐞𝐬, 𝐚𝐧𝐝 𝐫𝐞𝐟𝐢𝐧𝐞𝐬 𝐤𝐧𝐨𝐰𝐥𝐞𝐝𝐠𝐞. And if you are building AI systems in 2025, you will want to know the difference 👇 𝟏. 𝐇𝐲𝐛𝐫𝐢𝐝 𝐑𝐀𝐆 * Combines multiple retrieval methods (like search + vector DBs). * Think of it as layering multiple lenses to get richer, context-aware results. 𝟐. 𝐆𝐫𝐚𝐩𝐡 𝐑𝐀𝐆 * Organizes knowledge into graph structures. * This allows multi-agent workflows, memory, and parallel execution. * Imagine an AI agent that doesn’t just recall facts… but actually understands *relationships* between them. 𝟑. 𝐂𝐨𝐫𝐫𝐞𝐜𝐭𝐢𝐯𝐞 𝐑𝐀𝐆 * Detects errors in retrieved info and fixes them on the fly. * It is like spellcheck but for facts. * Ensures accuracy before the answer reaches the user. 💡 𝐖𝐡𝐲 𝐭𝐡𝐢𝐬 𝐦𝐚𝐭𝐭𝐞𝐫𝐬: The future of AI won’t just be about plugging a vector DB into an LLM. It will be about choosing the right RAG strategy for the problem you’re solving: * Rich context? → Hybrid * Complex relationships? → Graph * Accuracy under pressure? → Corrective 👉 Save this for later. 👉 Share with a friend building with RAG. Because in 2025, knowing the difference between RAGs could be the difference between a chatbot that sounds smart and an AI system that is actually trustworthy. #RAG #GraphRAG #CorrectiveRAG #HybridRAG #GenAI

Foundation Capital just published "Context Graphs: AI's Trillion-Dollar Opportunity" and it's the most technically coherent thesis I've seen on where enterprise AI infrastructure is heading. Let me break down why this matters. 🎯 The problem they identify is precise: when a renewal agent proposes a 20% discount despite a 10% policy cap, it needs to pull three SEV-1 incidents from PagerDuty, an open escalation from Zendesk, and a prior approval thread where a VP granted a similar exception last quarter. Today's CRM records "20% discount." All the reasoning vanishes. It's missing decision context. 📊 The architectural argument: organizational knowledge is inherently relational, not document-shaped. Sarah Chen belongs to Engineering Team. Engineering Team owns Payments Service. Payments Service has open incident INC-4521. INC-4521 was escalated by Customer Acme Corp. Acme Corp has Contract v3.2 with a 10% discount cap. When an agent needs to decide on that discount, it must traverse these connections. RAG finds documents containing "Acme" and "discount." Graph queries traverse actual relationships. That's the difference between similarity and meaning. 🏗️ The two-layer architecture emerging from this analysis: Layer 1 is operational context. Identity resolution so "Sarah Chen" and "S. Chen" and "sarah@company.com" resolve to one entity. Relationship modeling so ownership and accountability are queryable. Temporal state so you know what the contract said when the decision was made, not what it says now. Layer 2 is decision context. The traces capturing what inputs were considered, what policy version applied, what exception was granted, who approved it. This layer cannot exist without Layer 1. ⚙️ The technical implementation path involves compiling graph structure into constraints on generation. Build tries from the knowledge graph. Use them to mask invalid tokens during agent reasoning. Entity-constrained generation ensures agents only reference entities that exist. Path-constrained reasoning ensures agents only assert relationships that exist. Time-scoped trie construction ensures agents reason about historical state correctly. Property graphs win over RDF here. Cypher over SPARQL. Performance for traversals matters when agents need real-time context. But Schema.org vocabulary still valuable for the ontology layer. 🚫 Why incumbents cannot just add this capability: Salesforce stores current state. It knows what the opportunity looks like now, not what it looked like when the decision was made. Cannot replay decision-time state. Cannot audit, learn from, or use decisions as precedent. Snowflake sits in the read path. Receives data via ETL after decisions are made. By the time data lands, decision context is already gone. (Continue in comment)

MIT: Can AI Truly Evolve Its Own Knowledge Like Humans Do? Markus J. Buehler: Agentic Deep Graph Reasoning Yields Self-Organizing Knowledge Networks 👉 WHY THIS MATTERS  Most AI systems today learn from static data, producing answers in a single pass. But human knowledge grows through reflection, reconnecting ideas, and questioning past assumptions. What if AI could mimic this dynamic process, building and reorganizing knowledge iteratively? Markus Buehler’s latest research tackles this gap, proposing a system where AI doesn’t just retrieve information—it structures and refines it continuously, much like scientific discovery. 👉 WHAT THE PAPER SHOWS  The study introduces a framework where a language model and a graph network co-evolve. Instead of fixed data extraction, the AI actively generates concepts, links them into a knowledge graph, and uses the graph’s structure to guide its next queries. Over hundreds of cycles, this creates a network with two key traits: - Hubs: Highly connected nodes (e.g., core scientific principles) that anchor the system’s understanding. - Bridges: Concepts that link distant domains, sparking cross-disciplinary insights.    In materials science experiments, this approach led to discoveries that transcended simple data summaries, like bio-inspired designs for impact-resistant materials derived from biological repair mechanisms. 👉 HOW IT WORKS  The system operates through three recursive steps: 1. Generate: The AI proposes new concepts and relationships based on its current knowledge. 2. Merge: These ideas are added to a global graph, updating the network’s topology. 3. Refine: The graph’s structure informs the AI’s next line of inquiry, creating a feedback loop.    This mirrors how humans revisit hypotheses, but at scale: the graph grows without pre-defined rules, adapting as it integrates new information. 👉 The Bigger Picture  The research challenges the notion that AI must rely on static datasets or human-curated ontologies. By letting knowledge self-organize, Buehler’s work hints at a future where AI systems can autonomously explore scientific frontiers, draw unexpected connections, and propose solutions that blend disciplines—a step toward machines that think, not just calculate.

From ontology, knowledge graph, agents to workflows, a new paper demonstrates how a chemical-process digital twin can be built and used. Digital twins for chemical process have been discussed for over a decade, yet substantive detail rarely surfaces beyond vendor pitch decks. The gap fuels persistent confusion about where the technology actually sits today, and how ready it is for adoption. Published in Nature Chemical Engineering, Shuyuan Zhang et al. offer a concrete vision of such a twin, showing how it can be built on open standards and what it can deliver in practice. Here is how it is built: 🔹Ontology: Two paired schemas separate the universal physics of chemical processes from each plant's specific context, with rules deciding which laws apply where. 🔹Knowledge graph: Equations live in a machine-readable form alongside their original sources, with built-in links to chemistry databases and AI property predictors that supply data on demand. 🔹Agents: Specialized agents do the work, including assembling models from the graph, fitting parameters to experimental data, checking which laws apply, looking up properties, and answering chemistry questions through an LLM. 🔹Workflows: Two paths are supported. When the physics is known, the framework assembles a model bottom-up. When it isn't, it searches top-down across candidate laws, fits each in parallel, and selects the best fit. A useful reference for leaders and practitioners evaluating the digital-twin space, and a solid foundation to extend toward sensor integration, live calibration, and closed-loop control. 📄 A knowledge graph framework for digital twins of chemical processes, Nature Chemical Engineering, May 14, 2026 🔗 https://lnkd.in/eaq8DFSi