Vector Search, Graph Traversal, or Both? Choose by the Question

Use vector search when a question needs semantic similarity, graph traversal when it needs explicit relationships, and a combined approach when both are required. The choice should follow the evidence needed for an answer. A product does not need one retrieval method for every question.

Recognize the evidence shape

“Find incidents similar to this timeout” calls for semantic matching across descriptions. “Which services depend on this component?” requires reliable relationship data. “Which similar incidents affected services owned by this team?” needs both similarity and ownership constraints.

Write the expected answer and its supporting evidence before choosing an index. A relationship question cannot be solved reliably by hoping two names appear near each other in a vector result. A conceptual similarity question may not have a useful path through a graph at all.

Check whether the relationship is trustworthy

A graph is only as useful as its identities, edges, and update process. Two customer records with similar names should not be merged automatically. A dependency inferred from an old document should not silently override current configuration.

Record provenance and validity for relationships that affect decisions. Distinguish a source-stated relationship from an inference. If the data cannot support this distinction, a simpler retrieval answer that cites the source may be safer than a confident traversal result.

Combine methods at a clear boundary

One pattern finds candidate entities semantically, resolves their identities, then traverses permitted relationships. Another uses graph constraints to limit the document set before semantic ranking. Choose the order according to the query and available indexes.

Watch for a common failure: the first stage removes the only relevant evidence, so a sophisticated second stage has nothing useful to work with. Preserve stage-level traces and inspect candidate loss. Apply authorization throughout, including to expanded neighbors and supporting documents.

Evaluate three small question sets

Separate similarity questions, relationship questions, and mixed questions. Compare correctness, supporting evidence, empty-result behavior, latency, and operating work. Report each group separately so strong similarity performance cannot hide broken relationship answers.

Use the graph-RAG pilot guide to build the relationship test set. If you want managed memory for recurring agent questions, try Supermemory against those same questions and inspect its documented retrieval behavior instead of assuming every graph-based product exposes arbitrary traversal.