rag-knowledge-graph

Knowledge Graph RAG: Graph Traversal + Vector Search in Parallel

Vector search alone misses relationships. This pipeline extracts entities (Conductor/Software, Netflix/Organization), runs graph traversal and vector search in parallel, merges structured triples ({subject: "Conductor", predicate: "developed_by", object: "Netflix", confidence: 0.99}) with unstructured documents, and generates an enriched answer.

Workflow

question
    │
    ▼
┌────────────────────────┐
│ kg_extract_entities    │  NER: entities + types
└───────────┬────────────┘
            ▼
┌─── FORK_JOIN ──────────────────────────────────┐
│ ┌───────────────────────┐ ┌───────────────────┐│
│ │kg_graph_traverse      │ │kg_vector_search   ││
│ │(relationship triples) │ │(similarity search)││
│ └───────────────────────┘ └───────────────────┘│
└──────────────────────┬─────────────────────────┘
                       ▼
            ┌──────────────────┐
            │ kg_merge_context │  Combine facts + docs
            └────────┬─────────┘
                     ▼
            ┌──────────────────┐
            │ kg_generate      │  Generate from enriched context
            └──────────────────┘

Workers

ExtractEntitiesWorker (kg_extract_entities) -- Returns entities: {name: "Conductor", type: "Software"}, {name: "Netflix", type: "Organization"}.

GraphTraverseWorker (kg_graph_traverse) -- Returns triples: developed_by (0.99 confidence), is_a ("workflow orchestration engine", 0.98).

VectorSearchWorker (kg_vector_search) -- Returns 4 documents with entity hints and scores.

MergeContextWorker (kg_merge_context) -- Combines graph facts, relations, and vector docs. Handles null lists.

GenerateWorker (kg_generate) -- Calls gpt-4o-mini with the merged context.

Tests

34 tests cover entity extraction, graph traversal, vector search, context merging, and generation.

Further Reading

• RUNNING.md -- how to build and run this example