GraphNews

AI coding assistant skill (Claude Code, Codex, OpenCode, OpenClaw). Turn any folder of code, docs, papers, or images into a queryable knowledge graph - safishamsi/graphify

When video reasoning requires external knowledge, many systems with large multimodal models (LMMs) adopt retrieval augmentation to supply the missing context. Appending textual or multi-clip evidence, however, forces heterogeneous signals into a single attention space. We observe diluted attention and higher cognitive load even on non-long videos. The bottleneck is not only what to retrieve but how to represent and fuse external knowledge with the video backbone.We present Graph-to-Frame RAG (G2F-RAG), a training free and auditable paradigm that delivers knowledge in the visual space. On the offline stage, an agent builds a problem-agnostic video knowledge graph that integrates entities, events, spatial relations, and linked world knowledge. On the online stage, a hierarchical multi-agent controller decides whether external knowledge is needed, retrieves a minimal sufficient subgraph, and renders it as a single reasoning frame appended to the video. LMMs then perform joint reasoning in a unified visual domain. This design reduces cognitive load and leaves an explicit, inspectable evidence trail.G2F-RAG is plug-and-play across backbones and scales. It yields consistent gains on diverse public benchmarks, with larger improvements in knowledge-intensive settings. Ablations further confirm that knowledge representation and delivery matter. G2F-RAG reframes retrieval as visual space knowledge fusion for robust and interpretable video reasoning.

AI agents operating over extended time horizons accumulate experiences that serve multiple concurrent goals, and must often maintain conflicting interpretations of the same events. A concession during a client negotiation encodes as a ``trust-building investment'' for one strategic goal and a ``contractual liability'' for another. Current memory architectures assume a single correct encoding, or at best support multiple views over unified storage. We propose Rashomon Memory: an architecture where parallel goal-conditioned agents encode experiences according to their priorities and negotiate at query time through argumentation. Each perspective maintains its own ontology and knowledge graph. At retrieval, perspectives propose interpretations, critique each other's proposals using asymmetric domain knowledge, and Dung's argumentation semantics determines which proposals survive. The resulting attack graph is itself an explanation: it records which interpretation was selected, which alternatives were considered, and on what grounds they were rejected. We present a proof-of-concept showing that retrieval modes (selection, composition, conflict surfacing) emerge from attack graph topology, and that the conflict surfacing mode, where the system reports genuine disagreement rather than forcing resolution, lets decision-makers see the underlying interpretive conflict directly.

Retrieval-augmented generation (RAG) systems have been widely adopted in contemporary large language models (LLMs) due to their ability to improve generation quality while reducing the required input context length. In this work, we focus on RAG systems for long-document question answering. Current approaches suffer from a heavy reliance on LLM descriptions resulting in high resource consumption and latency, repetitive content across hierarchical levels, and hallucinations due to no or limited grounding in the source text. To improve both efficiency and factual accuracy through grounding, we propose GroundedKG-RAG, a RAG system in which the knowledge graph is explicitly extracted from and grounded in the source document. Specifically, we define nodes in GroundedKG as entities and actions, and edges as temporal or semantic relations, with each node and edge grounded in the original sentences. We construct GroundedKG from semantic role labeling (SRL) and abstract meaning representation (AMR) parses and then embed it for retrieval. During querying, we apply the same transformation to the query and retrieve the most relevant sentences from the grounded source text for question answering. We evaluate GroundedKG-RAG on examples from the NarrativeQA dataset and find that it performs on par with a state-of-the art proprietary long-context model at smaller cost and outperforms a competitive baseline. Additionally, our GroundedKG is interpretable and readable by humans, facilitating auditing of results and error analysis.

Knowledge graph construction typically relies either on predefined ontologies or on schema-free extraction. Ontology-driven pipelines enforce consistent typing but require costly schema design and maintenance, whereas schema-free methods often produce fragmented graphs with weak global organization, especially in long technical documents with dense, context-dependent information. We propose TRACE-KG (Text-dRiven schemA for Context-Enriched Knowledge Graphs), a multimodal framework that jointly constructs a context-enriched knowledge graph and an induced schema without assuming a predefined ontology. TRACE-KG captures conditional relations through structured qualifiers and organizes entities and relations using a data-driven schema that serves as a reusable semantic scaffold while preserving full traceability to the source evidence. Experiments show that TRACE-KG produces structurally coherent, traceable knowledge graphs and offers a practical alternative to both ontology-driven and schema-free construction pipelines.

Do You Really Need a Graph Database? How a Duck Can Serve Your Graphs You’re building a knowledge graph. Maybe it’s for an AI agent’s memory, a recommendation engine, or a semantic search …

Knowledge Management Strategy for the AI Era: Transforming Informational Chaos into Operational Performance Current generative AI is not just another software layer; it acts as a brutal revealer of …

This article introduces Graphify, an open-source tool that builds upon Andrej Karpathy's personal knowledge base concept to achieve fully multimodal automate...

𝐒𝐞𝐦𝐚𝐧𝐭𝐢𝐜 𝐋𝐚𝐲𝐞𝐫 𝐯𝐬 𝐒𝐞𝐦𝐚𝐧𝐭𝐢𝐜 𝐌𝐨𝐝𝐞𝐥𝐬: 𝐇𝐨𝐰 𝐓𝐡𝐞𝐲 𝐈𝐧𝐭𝐞𝐫𝐨𝐩𝐞𝐫𝐚𝐭𝐞 These two terms are used interchangeably often, not intentionally but often. They solve different problems and sit at different layers of the architecture. A semantic model is a declaration, while a semantic layer is a delivery mechanism. 🟡 𝐒𝐞𝐦𝐚𝐧𝐭𝐢𝐜 𝐌𝐨𝐝𝐞𝐥 The semantic model defines what exists in your domain: the entities, the relationships, the measures, the dimensions, and their precise business meaning. "Customer" means an entity with at least one completed transaction. "Revenue" means recognised revenue after refunds. 🟢 𝐒𝐞𝐦𝐚𝐧𝐭𝐢𝐜 𝐋𝐚𝐲𝐞𝐫 The semantic layer is what makes those commitments accessible. It's the governed interface through which downstream consumers (like BI tools, ML systems, APIs, and AI agents) query the model without reimplementing its logic. 🔐 𝐔𝐧𝐥𝐨𝐜𝐤𝐢𝐧𝐠 𝐒𝐞𝐦𝐚𝐧𝐭𝐢𝐜𝐬 In most organisations today, semantic definitions live inside tools. Power BI has its own. Tableau has its own semantics locked in. A Jupyter notebook has its own interpretation of defined and redefined entities. Alongside the inconsistencies, they are parallel contract violations, each one producing a slightly different version of truth for a different audience. A semantic model defined once and transpiled into each tool's native format eliminates this. Write once, generate everywhere. The model is the source of truth and the semantic layer is how you serve it to a broader audience with the right access protocols. Think high searchability, advanced searches, trail visibility, root cause lineage, cross-domain and cross-product visibility, and so on... 📑 𝐓𝐫𝐞𝐚𝐭 𝐭𝐡𝐞 𝐌𝐨𝐝𝐞𝐥 𝐚𝐬 𝐚 𝐂𝐨𝐧𝐭𝐫𝐚𝐜𝐭 But the interoperability only works when the model is designed with enough precision to survive translation. Vague semantic models produce vague semantic layers. If "revenue" is defined loosely in the model, the ambiguity compounds at every downstream tool that consumes it. This is why the semantic model isn't a BI configuration, but the 𝐬𝐞𝐦𝐚𝐧𝐭𝐢𝐜 𝐜𝐨𝐧𝐭𝐫𝐚𝐜𝐭 your entire consumption layer is written against. The layer without the model is a serving mechanism with nothing reliable to serve. #SemanticLayer #SemanticContracts #DataArchitecture

Karpathy's "second brain" is three folders and one schema file. raw/ holds source material, wiki/ holds AI-compiled articles, outputs/ holds generated answers. A CLAUDE·md defines the rules … topic structure, linking conventions, update triggers … and the AI maintains the wiki end-to-end. Each question enriches the next because answers get filed back. Monthly health checks catch the compounding errors that come with compounding knowledge. No database, no vector store, no Obsidian with 47 plugins … just nested .md files and a contract the AI can follow. Great hands-on guide on how to implement this linked in the comments. #knowledgeManagement #agenticAI #secondBrain #contextEngineering | 12 comments on LinkedIn

Graph validation in any language Bringing *translingual* capabilities to Apache TinkerPop is something I’ve been aiming for since a 2020 talk [1] — before Hydra existed. Some of the ideas from that talk (including “Escape from the JVM”) are now possible. In this demo, I show a common scenario: you have graph data, a schema, and you want to validate the data. What’s new is this: validation logic is no longer tied to a single programming language. You define validation once in any supported language, and it runs natively across languages with consistent semantics. This matters for TinkerPop and similar ecosystems, where multiple clients and components need to agree on shared logic. In other news, Hydra now has new implementations in Scala, Clojure, Common Lisp, and Scheme, with more on the way. Hydra: https://lnkd.in/g8v2hvd5 HydraPop demo: https://lnkd.in/g9p4iWPn Discord: https://bit.ly/lg-discord [1] (TinkerPop 2020): https://lnkd.in/gSC_cmbg

The 𝐌𝐨𝐝𝐞𝐫𝐧 𝐃𝐚𝐭𝐚 𝐑𝐞𝐩𝐨𝐫𝐭 2026 surveyed 540+ data leaders and professionals across 64 countries. The message that sticks: AI didn't create a new problem. It 𝐦𝐚𝐝𝐞 𝐚𝐧 𝐨𝐥𝐝 𝐩𝐫𝐨𝐛𝐥𝐞𝐦 𝐢𝐦𝐩𝐨𝐬𝐬𝐢𝐛𝐥𝐞 𝐭𝐨 𝐢𝐠𝐧𝐨𝐫𝐞. If we go back to basics, data exists to enable decisions. Everything else is 𝐢𝐧𝐟𝐫𝐚𝐬𝐭𝐫𝐮𝐜𝐭𝐮𝐫𝐞 𝐢𝐧 𝐬𝐞𝐫𝐯𝐢𝐜𝐞 𝐨𝐟 𝐭𝐡𝐚𝐭 𝐨𝐧𝐞 𝐠𝐨𝐚𝐥. By that measure, most data stacks are failing because they were never designed with the decision as the endpoint. The report refers to this phenomenon as the 𝘋𝘢𝘵𝘢 𝘈𝘤𝘵𝘪𝘷𝘢𝘵𝘪𝘰𝘯 𝘎𝘢𝘱. If we break it into its independent components, they are many problems cascading into one big gap - 🚨 𝐅𝐨𝐮𝐫 𝐝𝐢𝐬𝐭𝐢𝐧𝐜𝐭 𝐮𝐧𝐬𝐭𝐚𝐛𝐥𝐞 𝐭𝐫𝐚𝐜𝐤𝐬. 🔍 𝐃𝐚𝐭𝐚/𝐌𝐞𝐭𝐚 𝐃𝐢𝐬𝐜𝐨𝐯𝐞𝐫𝐲 89% say finding the right data is a top time drain. Data exists, but its location and relevance aren't encoded anywhere a human or machine can query reliably. 🕸️ 𝐂𝐨𝐧𝐭𝐞𝐱𝐭 𝐀𝐫𝐜𝐡𝐢𝐭𝐞𝐜𝐭𝐮𝐫𝐞 57% struggle to interpret data because 𝘮𝘦𝘢𝘯𝘪𝘯𝘨 is missing. 53% lack lineage. The definition has always lived in people's heads and has been considered the norm. 🧐 𝐓𝐫𝐮𝐬𝐭 𝐢𝐧 𝐌𝐞𝐭𝐫𝐢𝐜𝐬 93% encounter conflicting metrics. Nearly half don't fully trust their own data. When two dashboards show two numbers for the same metric, the ability to act collapses. 🔌 𝐃𝐚𝐭𝐚 𝐀𝐜𝐭𝐢𝐯𝐚𝐭𝐢𝐨𝐧 𝐰𝐢𝐭𝐡 𝐀𝐈 A third cannot trigger operational decisions using AI. Even when data is found and trusted by humans, it can't be consumed by machines without manual intervention. 𝐅𝐨𝐫 𝐲𝐞𝐚𝐫𝐬, 𝐡𝐮𝐦𝐚𝐧𝐬 𝐚𝐛𝐬𝐨𝐫𝐛𝐞𝐝 𝐭𝐡𝐞 𝐜𝐨𝐬𝐭 𝐨𝐟 𝐚𝐥𝐥 𝐟𝐨𝐮𝐫. An analyst would spend half a day finding the right table, confirm the metric definition with a colleague, and cross-reference dashboards before writing a single line of analysis. Inefficient, but it worked. 𝐀𝐈 𝐡𝐚𝐬 𝐧𝐨𝐧𝐞 𝐨𝐟 𝐭𝐡𝐨𝐬𝐞 𝐜𝐨𝐦𝐩𝐞𝐧𝐬𝐚𝐭𝐢𝐧𝐠 𝐦𝐞𝐜𝐡𝐚𝐧𝐢𝐬𝐦𝐬. It executes strictly on 𝘸𝘩𝘢𝘵 𝘵𝘩𝘦 𝘴𝘺𝘴𝘵𝘦𝘮 𝘮𝘢𝘬𝘦𝘴 𝘦𝘹𝘱𝘭𝘪𝘤𝘪𝘵. In most enterprises, very little is explicit. AI 𝘥𝘪𝘥𝘯'𝘵 𝘪𝘯𝘵𝘳𝘰𝘥𝘶𝘤𝘦 𝘯𝘦𝘸 𝘧𝘳𝘢𝘨𝘪𝘭𝘪𝘵𝘺. It removed the human layer that was absorbing it. ❎ A key signal: 80% ranked a semantic layer with standardised definitions as the top enabler of AI, 𝐚𝐛𝐨𝐯𝐞 𝐀𝐈 𝐭𝐨𝐨𝐥𝐬 𝐭𝐡𝐞𝐦𝐬𝐞𝐥𝐯𝐞𝐬, above faster processing. The report does 𝐍𝐎𝐓 describe a tooling gap - The MAD landscape is already falling over the edge, and this is not a problem for siloed or niche tools to solve. It is describing an 𝐚𝐫𝐜𝐡𝐢𝐭𝐞𝐜𝐭𝐮𝐫𝐚𝐥 𝐦𝐢𝐬𝐦𝐚𝐭𝐜𝐡 between the systems enterprises have built and the consumers those systems now need to serve. 🔖 That's the gap worth closing. Reference the 𝐟𝐮𝐥𝐥 𝐫𝐞𝐩𝐨𝐫𝐭: https://lnkd.in/gy8TGBrW

Over breakfast I spotted ‘Data Structures: Theory and Practice’ by Alfs Berztiss published 1971. Could not resist seeing what he had to say about graphs. “Let us now take a new approach ….”

Traverse, an embedded/server graph database similar to Neo4j and Memgraph that runs in production serving tens of millions of nodes and edges under high loads is now available in preview. Built from the ground up with a new query engine, it works across all operating systems with a single binary and has 99.95% openCypher compatibility and Bolt 5.x/6 support. Comes with a built-in IDE, gRPC, WebSockets, HTTP API, and MCP. Supports hot-swapping databases and TCP socket ingestion. It's tested against all known vendors and beats overall on all tested Pokec datasets. We wanted to take it a step further so we're including Silicon, a research unikernel build that runs enterprise software on Firecracker or Cloud Hypervisor and boots in under 100 ms. So fast it can boot platforms during a request. Traverse Silicon / Unikernel build is also available for preview. Silicon has a durable filesystem, SiliconFS, prioritizing crash safety. Unikernels provide full hardware isolation with no running OS, and the binary is under 10 MB. The attack surface is minimal since there is no operating system to exploit. Researchers, students, companies, and individuals can freely use the software, and given the simple embedded vs. server CLI architecture, AI builders can get started very quickly. Available preview: https://lnkd.in/dWvETFmp - Screenshot from console.

Is a "Personal Wiki" just a Knowledge Graph in disguise? 🤔 Andrej Karpathy’s idea of flat .md files "Second Brain" is incredibly popular right now. But we need to talk about the "Efficiency Gap." Karpathy’s solution: ✅ Readable ✅ Simple to start 📛 Inefficient at 10k+ nodes 📛 Massive token waste 📛 High latency for relationship-heavy queries If you have nodes and you have links, you have a Graph. Using an Agent to crawl flat files is a "brute force" approach to RAG. By moving to a GraphRAG architecture, you stop using tokens to "discover" relationships and start using them to "reason" over them. Elegance is great for a weekend project. Efficiency is required for a production-grade system. #AI #GraphRAG #Productivity #KnowledgeGraph

We all have systems for organizing things. Sometimes these systems are expressed outwards, like organizing a spice rack or a closet. Humans interact with organizational systems daily, with the likes of navigation systems, online shopping, the grocery store and even search engines.

This blog post shows you how to transform days of manual financial analysis into seconds of comprehensive insights by building an Agentic GraphRAG solution for capital markets. You’ll see the architecture, graph schema design, and agent setup that lets business users ask complex questions in plain language. Capital markets firms track financial relationships across multiple […]

If AGI means long-running autonomous tasks & multi-day agents, then persistent memory for agents is no longer optional. Memory is the new battleground in 2026. Especially after OpenClaw, I've seen a wild range of exploration, abstraction, and implementation around agent memory. Some go with complicated graph+vector+SQL retrieval, some just give up and rely on LLM+markdown+grep. I put together a presentation on agent memory, hope it helps people get started on this topic. | 19 comments on LinkedIn

A free white paper on AI taxonomy building. See practical methods for AI-assisted taxonomy creation for semantic search, knowledge graphs, and GraphRAG.

𝐖𝐡𝐞𝐫𝐞 𝐚𝐫𝐞 𝐭𝐡𝐞 𝐭𝐚𝐱𝐨𝐧𝐨𝐦𝐢𝐞𝐬 𝐢𝐧 𝐭𝐡𝐞 𝐬𝐞𝐦𝐚𝐧𝐭𝐢𝐜 𝐰𝐞𝐛 𝐬𝐭𝐚𝐜𝐤 I love the RDF “layer cake” as a helpful abstraction for understanding how our technologies build on one another. From URIs and RDF at the base up through ontologies, constraints, and reasoning. And yet, something essential seems to be missing: The everyday enterprise #vocabularies, #glossaries, and #taxonomies that organizations rely on to describe their world. Ontologies benefit enormously from well designed taxonomies as they bridge formally modeled knowledge with the language actually used inside a company. Conversely, we can start from the other direction: enriching the familiar enterprise vocabulary with more formal semantic structures. This creates a reinforcing loop in which terms evolve to classes, and relationships with greater clarity and consistency. A successful knowledge graph likely needs both worlds to some degree. The formal semantics of #RDFS, #OWL, and #SHACL give us powerful reasoning capabilities and data model flexible enough to represent our business. But the taxonomies and vocabularies provide the shared language that makes KG meaningful for real users and real business processes. 𝘚𝘰 𝘸𝘩𝘦𝘳𝘦, 𝘵𝘩𝘦𝘯, 𝘢𝘳𝘦 𝘵𝘢𝘹𝘰𝘯𝘰𝘮𝘪𝘦𝘴 𝘪𝘯 𝘰𝘶𝘳 𝘙𝘋𝘍 𝘴𝘵𝘢𝘤𝘬? The RDF stack focuses entirely on formalization and technical expressiveness. That focus makes sense when explaining inference, constraints, and data modeling. But it does not reflect how knowledge graphs are created and governed in organizations: Where taxonomy management, metadata stewardship, and terminology alignment are daily challenges. That gap is precisely why I enjoy conversations about interoperability and alignment matter so much. And it’s why I’m especially excited to be and speak at the upcoming Taxonomy Bootcamp conference in London in two weeks. (Interested in joining? I’ve got a speaker discount!). What’s your take on this? Does SKOS belong in the RDF stack? Or do we need another good abstraction on how the ontologies build on each other apart from formal semantics? (Thinking not only of SKOS but also PROV/DCTERMS/DCAT, etc.) | 17 comments on LinkedIn

Ontologies play a crucial role in healthcare systems due to the diversity of concepts, roles, users, and diagnostic and therapeutic methods. They facilitate the development of knowledge bases and the sharing and representation of information. With ...