Skip to content

Wearly Knowledge Graph

Learning objective. After working through this graph, you will be able to distinguish Wearly's three structured-knowledge layers (domain, user behavior, runtime), explain what node size encodes in terms of user signal, and describe how the graph functions as the substrate for future RAG and LLM shopping layers.

What you're looking at. A live render of graph/wearly-knowledge-graph.json — 91 nodes, 187 typed edges, generated deterministically from public seed data by scripts/generate_wearly_kg.py.

What this graph is for. It models the structured knowledge Wearly reasons over — reusable across users, queryable by future LLM layers. Big nodes = signals you produced more of; small nodes = declared in the domain but unused.

→ Open the Knowledge Graph in full screen ↗

Recommended for class demos. The full-screen view escapes the book's column width and gives the graph proper room to breathe.

How to read the graph

The viewer is built in the same design language as the SEIS 666 course learning-graph viewer: gradient page background, frosted-glass cards, soft pastel-per-type palette, simple toolbar.

Visual signal What it encodes
Color Node type. Each of the 17 types gets a distinct soft pastel; the User is the one dark anchor so the focal point is unmistakable. The legend shows every type with its count.
Shape Most nodes are ellipses (uniform, label-readable). The User is a star — single focal point. WardrobeGap and ShoppingSuggestion are triangles — warning shapes.
Size User signal — a wardrobe item worn more often + covering more occasions is larger; a color family used across many items is larger; a category with rich coverage is larger. Domain archetypes (rule packs, weather bands, palettes) stay at base size.
Edge color Light grey by default — un-emphasized. Highlighted edges (when a node is selected, search matches, or a legend type is filtered) turn periwinkle (#667eea).

What the controls do

Control Behavior
Search Type any name or type. Matching nodes stay full opacity; non-matches fade to ~18%.
Layer filter (sidebar) Filter to domain / user-behavior / runtime.
Layout dropdown (sidebar) Switch between Physics (force-directed) — default, lets categories cluster naturally — or Hierarchical (top-down).
Click any legend row Toggle to focus that node type. Everything else fades; the graph re-fits to the matching subset. Click the same row again to clear.
Click any node Open the info panel with type, layer, weight, one-sentence meaning, metrics, and in/out edge counts.
Double-click any node Highlight its 1-hop neighborhood; everything else fades to 15%. Click empty canvas to reset.
Fit to view Re-center and zoom to fit.
Reset Clear all filters; restore the default force-directed view.
Toggle Physics Pause / resume the simulation. Useful for capturing a screenshot at a stable layout.

The graph carries 20 typed relations. The full catalog:

Domain knowledge (static)

Edge Source Target Meaning
HAS_PROFILE User FitProfile The user's declared style + fit preferences.
HAS_SKIN_TONE User SkinTonePalette The user's color-palette anchor.
REQUIRES OccasionType WardrobeItemType The occasion's minimum required pieces.
USES_RULE OccasionType / WeatherCondition RulePack Which rule pack governs this context.
CONTAINS_RULE RulePack Rule Each pack's R-numbered rules.
EVALUATES Rule SkinTonePalette / WardrobeItemType Which axes a specific rule scores against.

User behavior (computed from the user's data)

Edge Source Target Meaning
OWNS User WardrobeItem The item is in the user's closet.
HAS_TYPE WardrobeItem WardrobeItemType Classification.
HAS_COLOR WardrobeItem / WishlistItem ColorFamily Color bucket.
SUITABLE_FOR WardrobeItem OccasionType Has an occasion tag.
HAS_FAVORITE_STORE User FavoriteStore Saved retailer.
HAS_WISHLIST_ITEM User WishlistItem Wants to acquire.
PREFERS User ColorFamily Aggregate preference signal.

Runtime archetypes (the bridge to live recommendations)

Edge Source Target Meaning
POWERS RulePack WorkflowStep The rule pack runs at this step.
PRODUCES WorkflowStep OutfitRecommendation / WardrobeGap / ShoppingSuggestion Each step's output archetype.
CONTAINS_ITEM OutfitRecommendation WardrobeItem Items in an outfit.
FLAGS_GAP OutfitRecommendation WardrobeGap Per-run gaps.
SUGGESTS WardrobeGap ShoppingSuggestion Gap-driven suggestion.
REJECTS Feedback WardrobeItem User pushback.
WISHLIST_CLOSES_GAP WishlistItem WardrobeGap A wished item that would close a current gap.

How node size is computed

Node type Size formula
WardrobeItem 1.0 + 0.20 × worn_count + 0.10 × versatility (versatility = count of distinct canonical occasion tags). A piece worn six times across three occasions reaches ~2.4.
WardrobeItemType, ColorFamily, OccasionType Rolled up from connected WardrobeItem weights. A category with many heavy items becomes a heavy category node.
FavoriteStore Scales with times_chosen (when overlay data is included).
All other types Base size — domain/runtime archetypes are equal-importance facts.

All weights clamp to [1.0, 3.0] so a heavy category never dwarfs the User node.

How this enables RAG and an LLM shopping agent

The graph is structured context — the kind of input modern LLM agents perform better against than a flat-text dump of the same information.

RAG over the Intelligent Book

A question like "Why does this dress work for my profile?" becomes:

  1. Start at the dress's WardrobeItem node.
  2. Walk one hop to WardrobeItemType + ColorFamily + SUITABLE_FOR occasion(s).
  3. Follow USES_RULE to the rule packs that govern those contexts.
  4. Retrieve the book chapters keyed off the rule slugs the agent already emits in its reasoning trail.

The LLM gets a compact subgraph + the relevant book passages. Retrieval happens along graph edges, not vector embeddings, so the explanation cites the same rule the runtime agent does.

LLM shopping agent over favorite stores

When a WardrobeGap is detected at runtime, the LLM shopping agent:

  1. Reads the gap's connected WardrobeItemType and OccasionType.
  2. Walks the user's HAS_SKIN_TONE for the target palette and HAS_PROFILE for the silhouette.
  3. Visits each FavoriteStore and queries that store's catalog with the graph-derived constraints.
  4. Scores any candidate against the same rule packs the runtime engine uses.

The rule engine stays the safety + scoring substrate; the LLM is the discovery layer.

Structure beats prompts

  • Compact — the full graph is ~66 KB; the equivalent flat-text description is an order of magnitude larger.
  • Queryable — filtering by layer / type / metric returns a precise subgraph.
  • Auditable — every edge has a typed from / to / type. The citation chain is literally a graph traversal.

Linked concept

See also