GraphNews

White Paper: "Semantic Foundations for AI-Ready Data: Why Governed Semantics, Open Standards, and Data Products Are the Prerequisites for Enterprise AI.," Synthesized with Claude Opus 4.6 (with guardrails) from Gartner D&A Summit 2026 proceedings, Gartner research publications, and industry analysis.

Retrieval-Augmented Generation (RAG) systems typically treat documents as flat text, ignoring the structured metadata and linked relationships that knowledge graphs provide. In this paper, we...

🧩 What Are Modalities in Knowledge Graphs — And Why They Matter

🧠 This one's flying under the radar and it shouldn't be.

Kinetica-Graph Explorer is now available.

Competency Questions for System Architectures

The shift from "finding data" to "reasoning and understanding context" is the driver for a more robust context layer that needs to provide the operational nuances that are typically locked within systems, teams, and organizational silos.

There are essentially five potential vectors toward shared understanding of semantics: * #ad_hoc. This is the JSON model; send the data without schemas, let the receiver figure it out. Usually involves human-in-the-loop and much swearing when things don't work as agreed upon. * #contractual . You and I agree a priori on a common schema or ontology. This can be extended to a certain extent to a network of users sharing that same schema. This is the semantic web model, and it works reasonably well, but also means that communicating nuance or variation is hard, especially the larger the network becomes. * #declarative . When I transmit my data, the ontology is part of that data. The recipient of the data then has instructions concerning how to work with that data. This increases the overall size of data transmissions, but has the advantage of packaging data and schema together so that the recipient does not necessarily need point-to-point synchronous communication. This is the data package approach, and works very well with SHACL. Its principal disadvantage is that while schematic information is usually fairly condensed, taxonomic information can be extensive. * #negotiative. A system receives a data packet and attempts to ascertain its schematic rules contextually, in terms of a known internal schema resolution (or a derived schema from some other data context). Negotiative schemas have the advantage in that they don't require complex parsers and transformation layers (though you can argue that an LLM or similar cognitive system is hardly a trivial intermediation layer) but they suffer from ambiguity, lack of provenance, and don't deal with large, complex data structures well. * #cached_transformative. A variation of #negotiative, the participants in a data exchange each submit target schemas, which are then used by a transformer (perhaps an LLM) to negotiate some form of point-to-point transformation from one schema to the next, with each schema identified by a unique IRI. The transformation can work with test data to ascertain degrees of lossiness (some lossiness is expected), then, when this drops below an acceptable threshold (or can be filtered out via validation), the transformations are cached. Data so transmitted then includes the source and target schema IRIs to retrieve the relevant transformation. Most data systems use some or even all of these approaches. JSON is mostly #ad_hoc, JSON-LD is somewhat #declarative. Most RDF is #contractual - you work against existing schemas. #negotiative and #cached_transformative are both comparatively recent, and are dependent upon intelligent agents; the primary difference is that #negotiative transformations are repeated for every transaction, #cached_transformative are only regenerated when a schematic change takes place. Each has its place.

The origin of most data in the world comes from closed-world systems. Forms, logging systems, and sensors all generate data in predefined structures where the schema is known in advance. Yet when peop

Google Gemini has totally crushed all other models according to Knwler's benchmark.

Table of Contents

Most Graph RAG discussions are still around 𝘸𝘩𝘪𝘤𝘩 𝘨𝘳𝘢𝘱𝘩 𝘥𝘢𝘵𝘢𝘣𝘴𝘦 𝘵𝘰 𝘶𝘴𝘦? Serious architecture asks: Which graph structure fits the retrieval job? That is the Graph-Type Fit Pattern. The mistake is thinking “graph” is a feature.

Gartner D&A is always a mix of signal and noise.

A daily-refreshed catalog of knowledge graph resources built on Wikidata

Modern data architectures are fragmented across graph databases, vector stores, analytics engines, and optimization solvers, resulting in complex ETL pipelines and synchronization overhead. We...

Retrieval-Augmented Generation (RAG) systems combine Large Language Models (LLMs) with external knowledge, and their performance depends heavily on how that knowledge is represented. This study...

In this post, I want to draw attention to what I think is a very important, yet under-appreciated, aspect of the Semantic Web. Let's start with a brief intro to a latter, for whom is not familiar with it.

Browse ontologies, controlled vocabularies, taxonomies, and semantic software from a daily-refreshed RDF catalog.

Over the past several years I’ve been thinking about how AI systems will actually operate inside real enterprises. Most discussions focus on models, pipelines, or agents. Those things matter, but they only describe parts of the system. Enterprises don’t operate on probability alone. They operate within defined objectives, authority structures, and governance processes. That raises a different architectural question. How should probabilistic AI reasoning operate inside institutions that ultimately make decisions through governed processes? The article below outlines a framework I’ve been developing called the AI Knowledge Framework (AIKF). It connects five parts of enterprise AI architecture that are often discussed separately: - Purpose - Meaning - Evidence - Reasoning - Governance The central idea is straightforward. AI systems are good at discovering patterns and generating evidence. Institutions remain responsible for determining meaning, responsibility, and consequences. Purpose defines the boundaries within which AI reasoning operates. I’d be interested in hearing how others working on enterprise ontologies, knowledge graphs, semantic architectures, or governed AI systems are thinking about these issues. | 43 comments on LinkedIn

Visualise any GitHub repo as a graph. Simply replace "github" with "gitcgr" in the repo URL. This gives you an interactive graph of the entire codebase structure, with functions, classes, modules, and their relationships all laid out visually. Internally, it's powered by Code-Graph-RAG, an open-source system that does deep code analysis under the hood: - It uses Tree-sitter to parse your codebase across multiple languages. - Builds a full knowledge graph of your repo in Memgraph. - Lets you query your codebase in natural language. Code-Graph-RAG also works as an MCP server, which means you can plug it directly into Claude Code and query or edit your codebase using natural language from your terminal. ___ Interested in ML/AI Engineering? I have put together a free PDF (380+ pages) with 150+ core lessons. Find it in the comments! ___ Share this with your network if you found this insightful ♻️ Find me → Avi Chawla. Every day, I share tutorials and insights on DS, ML, LLMs, and RAGs. | 10 comments on LinkedIn

Is Palantir onto something, or are they just blowing hot air? You may have heard the term Ontology, and why leading businesses rely on them to build trustworthy data models. Recent advances in LLMs have made it possible for startups and enterprises alike to build flexible, reliable semantic layers faster

Graphs are a powerful way to model complex relationships, from social networks to supply chains.

People are conflating data context and decision context.

From Berlin to San Francisco On the same day, two technical communities one in Berlin and one in San Francisco gathered to discuss the same architectural question: how should memory be designed for agentic AI systems? In Berlin, AI Memory and… Read more →