Where time is spent: seed search (vector similarity) vs full retrieval (search + BFS graph walk). The gap shows BFS overhead — ideally minimal.
How much the seed results overlap with BFS results. Lower overlap = BFS is finding more unique nodes the vector search missed. This proves the graph structure adds value beyond flat retrieval.
End-to-end retrieval time per query. First query is slow (embedding model cold start), subsequent queries are fast. Should stay constant regardless of graph size — that's the O(log N) guarantee.
| Time | Type | Duration (ms) | Details |
|---|---|---|---|
| Loading... | |||