Neo4j GraphRAG

Build a typed incident graph when users need to ask both semantic and relationship questions, such as which incidents affect a product and what source chunks support that connection.

Pattern

• Define strict Product and Incident nodes plus an AFFECTS relationship.
• Chunk incident documents with stable document and chunk IDs.
• Extract graph facts with a completion model.
• Embed chunks and canonical entity descriptions with the same vector dimensions as the graph indexes.
• Replace affected documents transactionally.
• Give the agent a bounded hybrid graph-search tool with chunk evidence.

import { createGraphSearchTool } from '@anvia/graph'

const searchGraph = createGraphSearchTool({
  name: 'search_incident_graph',
  description: 'Search incidents, affected products, and supporting evidence.',
  graph,
  model: embeddingModel,
  search: {
    type: 'hybrid',
    seeds: ['chunks', 'entities'],
    topK: 6,
    candidatesPerSeed: 12,
    rrfK: 60,
  },
  traversal: {
    relationships: ['AFFECTS'],
    direction: 'both',
    maxDepth: 2,
    maxNodes: 24,
    maxRelationships: 36,
  },
  evidence: { type: 'chunks', maxChunks: 8 },
})

Ask which incidents affect a known product, then ask for the supporting source text. Validate the returned entity and relationship identities plus eligible evidence IDs instead of asserting exact model prose.

Failure cases include index-dimension mismatches, extraction schema violations, conflicting entity properties, unknown provenance chunk IDs, unavailable full-text indexes, and traversal limits that are too small for the required path.

See the runnable v1 cookbook source.