@invana/graph-datasets

Ready-to-render example graphs for @invana/graph — each one engine-ready data plus the settings that make it look right.

pnpm add @invana/graph-datasets

A dataset is a folder

Every dataset lives in src/<name>/ as a single data.ts exporting two halves — data (what to draw, as @invana/graph's GraphData) and settings (how it should look, as @invana/canvas's CanvasConfig) — so a complete visualisation is two props and no glue:

import { lesMiserables, lesMiserablesSettings } from '@invana/graph-datasets';

<GraphCanvasApp data={lesMiserables} config={lesMiserablesSettings} />

Moving to one subpath per dataset (@invana/graph-datasets/les-miserables → { data, settings }) — see docs/graph-datasets-folder-exports-plan.md. The folder names below are stable across that change.

The datasets

24 sets, largest first. Bold is the layout settings.activeLayout selects; the rest are layouts the data supports. † = under @invana/graph-datasets/usecase-demos; ‡ = its own subpath (too large for the main bundle). Counts are the shipped value — generated sets show their default parameters.

| Dataset | Kind | Nodes | Edges | Layouts | What it is | |---|---|---:|---:|---|---| | h1b2019 | hierarchy | 24 683 | 24 682 | hierarchy (radial) · pack · force | H-1B 2019 state → city → employer. Four levels, thousands of leaves. | | game-of-thrones ‡ | network | 4 959 | 28 679 | force · geometric | Houses · characters · locations · seasons. The scale test. | | air-routes | geo | 2 980 | 0 | none — a map projection places them | World airports (lng/lat on data) + a land TopoJSON basemap. | | cora † | network | 2 708 | 10 556 | force · geometric | The Cora citation benchmark; type is the paper's subject area. | | wikipedia-dataviz ‡ | network | 2 085 | 5 409 | none — positions are the data · force | Sigma.js data-viz knowledge map; 24 topic clusters + its own meta-graph. | | invana-code-kg † | network | 602 | 1 329 | force · elk | This repo as a code knowledge graph — files · functions · classes, plus clusters, a guided tour and provenance. | | rag-embeddings † | point cloud | 402 | 0 | none — positions are the data | A 2-D embedding projection over five topics. The clustering is the picture. | | lattice | network | 400 | 760 | force · geometric | A 20×20 grid. Regular structure to sanity-check a layout against. Generated. | | life-tree | hierarchy | 381 | 380 | hierarchy (radial) · tree · force | Tree of life parsed from Newick; branch lengths on data.length. | | old-faithful | point cloud | 272 | 0 | none — positions are the data | Eruption duration vs waiting time. A scatter plot rendered as a graph. | | flare-imports | hierarchy | 252 | 906 | force · hierarchy (radial) | Flare plus synthetic leaf→leaf imports, for edge bundling. Tree and import edges are kept apart. | | flare | hierarchy | 252 | 251 | hierarchy (tree) · radial · cluster · pack · force | The Flare package tree — the canonical d3-hierarchy fixture. Also ships the nested tree. | | citations † | network | 150 | 402 | force · geometric | Synthetic paper citations across five ML topics. | | random-tree | hierarchy | 120 | 119 | force · hierarchy (tree) | A generated random tree — throwaway structure at any size. | | twitter | network | 99 | 196 | force · elk | Synthetic activity graph — users · tweets · comments · retweets · hashtags. Generated. | | les-miserables | network | 77 | 254 | force · elk · geometric | Hugo character co-occurrence (Knuth, Stanford GraphBase). The small dense classic. | | uk-energy-flow | flow | 48 | 68 | sankey | UK energy supply → conversion → end use. Ribbon width is the datum. | | invana-architecture † | diagram | 43 | 33 | none — hand-placed | The Invana pipeline, hand-placed with parentId containment. | | ontology † | network | 29 | 36 | force · elk | A business ontology — companies · people · products · locations. | | microservices † | network | 20 | 34 | force · elk | A service topology with tier, health and RPS per node. | | computing-pioneers † | network | 10 | 7 | force · elk | Ten pioneers with avatars — the composite-card fixture. | | agent-trace † | DAG | 9 | 9 | elk (layered, down) · hierarchy (tree) | An LLM agent run: prompt → tools → decision → response. Three presets; the first is the default. | | star-schema † | schema | 4 | 3 | elk (layered, right) | A fact table and its dimensions, as ER cards with typed fields. | | modeller-seed † | diagram | 3 | 1 | none — hand-placed | Three seeded nodes — the empty-canvas starting point for the modeller. |

Kinds. network — structure is the data, geometry comes from a layout. hierarchy — a rooted tree flattened to parent→child edges (the nested tree ships too). flow — directed magnitudes; edge weight is the point. DAG — a directed run/trace; ordering is the data. schema — a meta-graph of entity kinds, not instances. point cloud / geo / diagram — the node positions are the information, so activeLayout is ''; running a solver over one destroys the picture.

Mounting a layout. settings.activeLayout names an id the consumer must have mounted. graph-force ships in the <GraphCanvasApp> bundle, so force sets drop straight in. Hierarchy and Sankey sets use the id layout (mount D3HierarchyLayout — the mode is already in the settings — or D3SankeyLayout); agent-trace uses the id elk (mount ElkLayout).

Adding a dataset

See CLAUDE.md — the authoring contract (engine-ready records, type as the discriminator, offline transforms in scripts/, settings as pure JSON) and the four steps.