A question that needs both
“Which current support policies apply to products that depend on the billing service changed last week?” needs a graph to follow product and service relationships, a date filter to identify the change, and document retrieval to read the actual policies. Neither vector search nor a graph alone is enough.
How the two layers divide the work
- The graph identifies entities, relationships, filters, and paths.
- Vector search finds supporting passages, explanations, and unstructured evidence.
- The application joins both results and shows where each claim came from.
- A reviewer can inspect the path and the source text instead of trusting a blended score.
Use a simpler system when
A document collection can answer the questions with ordinary RAG, or the relationships are already cleanly represented in SQL. Hybrid architecture adds ingestion, synchronization, query planning, and validation work. Use it when the relationship queries are frequent and valuable enough to justify that cost.
The first proof
Take 20 real multi-step questions and label the expected entities, path, and supporting passages. Compare a document-only baseline with the hybrid design. If the hybrid system cannot improve answer completeness or traceability on those questions, do not keep it because the diagram looks sophisticated.