router-core: DistanceMetric::DotProduct is inverted — HNSW search returns the LEAST similar vectors

[ruvnet]

Summary

ruvector-router-core's DistanceMetric::DotProduct returns the raw inner product as a distance (src/distance.rs, the DotProduct arm), while index.rs keeps the smallest distances during both graph construction and search. The result: with DotProduct, search() ranks the least-similar vectors first, and the graph itself is built with neighbours chosen by minimum inner product, so the index is not navigable for maximum-inner-product search.

Evidence

While building ruvector-kge (branch feat/kge-package) we measured recall@10 of DotProduct HNSW vs exhaustive maximum-inner-product on 2,000 random 64-d vectors: 0.000. Negating the query does not help, because the entity–entity graph edges are also oriented by minimum dot product. The measurement is preserved as an #[ignore] test, raw_dotproduct_recall_for_the_record, in crates/ruvector-kge/src/ann.rs on that branch.

The same setup using the standard MIPS→L2 reduction (append sqrt(phi^2 - ||x||^2) to each indexed vector, append 0 to the query, DistanceMetric::Euclidean) gives recall@10 0.996, which is what ruvector-kge ships as its workaround.

Expected

Either DotProduct distance = -dot (or 1 - dot for normalised vectors) so that smaller-is-closer holds everywhere the index compares distances, or document DotProduct as unsupported for HNSW and reject it at VectorIndex::new.

Notes

• Cosine is unaffected (it already returns a distance).
• Related: the neighbour-selection heuristic in PR fix(router,ruvector): ship #430 + HNSW neighbour-selection heuristic; honour documented DbOptions; router 0.1.31, ruvector 0.3.2 #1005 assumes smaller-is-closer, which is another reason the metric must be a true distance.

🤖 Generated with claude-flow

https://claude.ai/code/session_01LNEfix9xQvKBUcVVSBCWXa

[ruvnet]

Investigated in #1014: router-core's DotProduct is not inverted. distance::dot_product has returned -(a·b) since Nov 2025, so the index's smallest-distance-first search already is a maximum-inner-product search.

The recall@10 = 0.000 came from the ruvector-kge diagnostic negating the query again, which searches for the minimum inner product. With the query passed through unchanged, the same 2,000 × 64-d table measures recall@10 = 1.000.

#1014 turns that ignored diagnostic into a gate (dotproduct_recall_matches_mips) and corrects the kge module docs. router-core is unchanged. The kge L2 reduction remains valid, so switching kge to plain DotProduct is optional.