The Problem
HNSW picks a random entry point and greedily descends. For graph traversals, a random start means more hops to reach the target. And B-tree index lookups for every neighbor of a hot node waste cycles on cache misses.
PageRank Entry Points
PageRank identifies nodes that many paths flow through. Starting traversals at high-PageRank nodes means fewer hops to reach any destination — the same principle that makes Google's web search start from authoritative pages.
Traversal Counters
As queries run, sutraDB increments a counter each time a node's neighbors are looked up. When a node crosses a hotness threshold, its adjacency list is materialized — all neighbors pre-loaded into contiguous memory. This is adaptive: cold nodes stay in the B-tree, hot nodes get fast flat arrays.
Steps