Question answering over a graph

Goal: answer questions like “Who on Alice’s team works in Seattle?” that need connected facts.

Pattern: use read() to find the anchor, traverseGraph() to expand it, and inject both into the prompt.

import { formatContext, formatGraphContext } from '@equationalapplications/core-llm-wiki';

export async function answer(wiki, entityId: string, question: string) {
  // 1. Semantic search finds the most relevant starting fact
  const hits = await wiki.read(entityId, question, { maxResults: 5 });
  if (hits.facts.length === 0) return generate(question, '');

  // 2. Expand the top hit into its neighbourhood
  const graph = await wiki.traverseGraph(entityId, {
    sourceId: hits.facts[0].id,
    maxDepth: 2,
    direction: 'both',
    minTraversalConfidence: 'inferred',
    maxTraversalNodes: 25,
    excludeDrafts: true,
  });

  // 3. Give the model both views
  const context = [
    '## Relevant facts', formatContext(hits, { maxTasks: 0, maxEvents: 0 }),
    '## Connected facts', formatGraphContext(graph),
  ].join('\n\n');

  return generate(question, context);
}

Prerequisites

The graph only has edges if an ontology was active when facts were written. The quickest setup:

import { schemaOrgWarmAgentManifest } from '@equationalapplications/schema-org-llm-wiki';
await wiki.setOntologyManifest(entityId, schemaOrgWarmAgentManifest, { mode: 'strict' });
await wiki.runOntologyBackfill(entityId, { classifier: 'llm' }); // types older facts + proposes edges

Tips

• Filter with edgeTypes (for example ['worksFor', 'memberOf']) for focused questions.
• Run lint(entityId) occasionally. A high danglingEdges count suggests stale references.