@equationalapplications/schema-org-llm-wiki

Curated schema.org warm-agent ontology manifest for hybrid LLM memory. Seeds a knowledge graph with 9 standard node types and 28 polymorphic edges — token-efficient, JSON-LD-ready, data-only with no runtime code.

📖 Docs · GitHub · ScopeLab · WikiDemo · Changelog · Issues

Ontology manifest for Core LLM Wiki, inspired by Andrej Karpathy's LLM Wiki memory spec.

Why curated?

Full schema.org has ~800 types and ~1,400 properties. Injecting that into every librarian/ingest prompt would blow token budgets and collapse LLM classification accuracy. This manifest selects the high-value warm-agent subset (~2 KB serialized, 9-way classification) while keeping every type and property standard, so stored facts and edges map 1:1 onto schema.org for future JSON-LD export.

Why this is the GraphRAG ontology

This manifest is designed as the canonical taxonomy bundle for warm-agent GraphRAG: the 9 node types and 28 polymorphic edges cover the people, places, organizations, projects, events, and creative works that dominate personal and professional knowledge graphs. Pair it with core-llm-wiki's traverseGraph() and you get a SQL-only GraphRAG stack with no Neo4j.

Why a curated ontology prevents hallucination in edge extraction

The librarian and ingest LLM passes that write llm_wiki_edges only see this manifest's node types and edge properties in their prompt context. Without that constraint, an unconstrained LLM will invent edge types and node types on every call — producing a noisy graph where WITH RECURSIVE walks return orphaned nodes and arbitrary relationships.

With the manifest:

• Every edge has a valid (type, source_type, target_type) triple — the manifest validates edge structure and reduces invalid relationships.
• Polymorphic edges (location, organizer, about, itemReviewed) cover the cases where a single property name applies to many source/target type combinations.
• Token budget stays small — ~2 KB serialized, vs ~50 KB for the full schema.org catalog.

Use it with core-llm-wiki for GraphRAG

import { createWiki } from '@equationalapplications/core-llm-wiki';
import { schemaOrgWarmAgentManifest } from '@equationalapplications/schema-org-llm-wiki';

const wiki = createWiki(db, {
  llmProvider,
  config: {
    ontology: {
      mode: 'strict',
      seedManifests: {
        [entityId]: { mode: 'strict', manifest: schemaOrgWarmAgentManifest },
      },
    },
  },
});

// After ingestDocument / runLibrarian populate edges:
const graph = await wiki.traverseGraph(entityId, { sourceId, maxDepth: 2 });

Requirements

Requires @equationalapplications/core-llm-wiki at the same release or newer — this manifest uses polymorphic edge rows (one property name with several source/target types), which core validates by the (type, source_type, target_type) triple.

Installation

npm install @equationalapplications/schema-org-llm-wiki

Usage

import { createWiki } from '@equationalapplications/core-llm-wiki';
import { schemaOrgWarmAgentManifest } from '@equationalapplications/schema-org-llm-wiki';

const wiki = createWiki(db, {
  llmProvider,
  config: {
    ontology: {
      mode: 'strict',
      seedManifests: {
        [entityId]: { mode: 'strict', manifest: schemaOrgWarmAgentManifest },
      },
    },
  },
});

Or seed an existing entity directly:

await wiki.setOntologyManifest(entityId, schemaOrgWarmAgentManifest, { mode: 'strict' });

Node types

Edge properties (19 names, 28 rows)

All standard schema.org properties: knows, spouse, parent, worksFor, memberOf, homeLocation, workLocation, location (×2), containedInPlace, subOrganization, object, agent, attendee, organizer (×2), author, publisher, about (×4), itemReviewed (×5), owns.

Polymorphic properties appear as multiple rows with distinct source/target types, mirroring schema.org's own domain/range definitions — e.g. about targets person, organization, place, and event.

JSON-LD export notes

• Property names keep schema.org camelCase (worksFor, itemReviewed); core validation compares case-insensitively but stores the given casing, so names survive to storage and future JSON-LD export unchanged.
• Literal-valued properties (birthDate, startTime, reviewRating values, …) live inside fact content, not as edges. Only object-valued properties (pointing at other facts) are edges.

Monorepo Ecosystem

License

MIT

Made with ❤️ by Equational Applications LLC. https://equationalapplications.com/