Structure Plus Content, Not Structure Instead Of Content

The Principle

Knowledge graphs for RAG should integrate textual chunks with symbolic structure, not replace text with triples. The graph provides navigability; the chunks provide the actual content needed for generation.

Why This Matters

There’s a tempting assumption that graph-based RAG means converting everything into entities and relationships. But when researchers tested ontology-guided knowledge graphs without integrated text chunks, accuracy dropped from 90% to 15-20%.

The symbolic structure tells you where to look. The text tells you what to say.

How to Apply

When building knowledge graphs for RAG:

1. Extract entities and relationships to create the navigational structure
2. Attach source chunks to graph nodes so the original text remains accessible
3. Retrieve subgraphs with their associated chunks when answering queries
4. Let the LLM generate from the chunks, using the graph structure to ensure relevant content is included

The graph is the index. The chunks are the content.

When This Especially Matters

• Domain-specific Q&A where accuracy is critical
• Multi-hop reasoning that requires connecting information from multiple sources
• Any RAG implementation where you’re considering “pure” graph representations

Exceptions

For very structured domains with complete ontologies (e.g., certain biomedical applications), entity-relationship triples alone may suffice. But for most enterprise knowledge, the original text carries nuance that triples can’t capture.

Related: 07-molecule—vectors-vs-graphs, 06-atom—ontology-guided-kg-construction, 07-molecule—rag-core-tradeoffs