A headline says "Arm unveils AGI CPU." A knowledge graph extracts the structure: Arm → launch → AGI CPU. Another headline says "Meta and OpenAI are among the first customers." The graph adds: Arm → partnership → Meta, OpenAI. A third says "Nvidia cut its Arm stake by 43.8%." The graph records: Nvidia → financial → Arm (divestment). Three headlines become three edges. The edges form a network. The network reveals a story no single headline tells: the architect became a competitor.
How TEXXR Builds Its Graph
The knowledge graph is built in three stages:
1. Entity extraction. NLP models scan every article and identify entities — companies, people, products, technologies. "Apple" is tagged as a company. "Tim Cook" as a person. "iPhone" as a product. "Core ML" as a technology. This turns unstructured text into a searchable entity database.
2. Relationship extraction. An LLM reads each headline and description and identifies structured relationships between entities. Each relationship has subjects (who), a predicate (what they did), objects (to whom), a summary (human-readable description), and a status (confirmed, rumored, denied, failed).
3. Categorization. Each relationship is assigned a canonical category: partnership, acquisition, launch, financial, controversy, legal, regulatory, personnel, or competitor. This standardization allows comparison across time — you can track how an entity's relationship profile shifts from "mostly partnerships" to "mostly competition."
What Edges Look Like
Each edge in the graph contains:
| Field | Example |
|---|---|
| Subjects | OpenAI |
| Predicate | partnership |
| Objects | Microsoft, Azure |
| Summary | OpenAI expands its exclusive cloud partnership with Microsoft Azure |
| Status | confirmed |
| First seen | 2023-01-23 |
| Last seen | 2025-11-18 |
| Articles | 12 source articles |
The first_seen and last_seen dates make edges temporal — you can see when a relationship started and when it was last mentioned. An edge with a wide date range is an enduring relationship. One that appears and disappears quickly was a brief event.
What the Graph Reveals
Predicate evolution. TEXXR's edge-arc tool tracks how an entity's relationship profile changes over quarters. When Arm's profile shifted from 36% "launch" + 36% "financial" (2024) to 50% "competitor" + 50% "controversy" (2026), the knowledge graph was registering a structural identity change — months before it became obvious.
Relationship threads. Following the edges between two entities over time tells a story. The Arm-Nvidia thread showed: joint ventures (2024) → Nvidia cutting its stake (Feb 2025) → continued partnership on NVLink (Nov 2025) → direct competition (Mar 2026). Each edge was a chapter. The thread was the book.
Status flows. Edges have status: confirmed, rumored, denied, failed. Tracking status changes reveals how stories evolve — a rumored acquisition that gets confirmed, a partnership that fails, a product launch that gets denied. The Nexus interface lets you filter by status to see the lifecycle of a story.
Try It
The Nexus page lets you explore the knowledge graph directly — search for any entity and see its relationships. The Pulse API provides programmatic access via /api/hyperedges/entity/OpenAI and /api/hyperedges/between/Arm/Nvidia.
For a deeper dive into how the graph powers editorial analysis, see How TEXXR Works.