research(nightly): multi-subspace HNSW with coherence-weighted fusion#556
Draft
ruvnet wants to merge 1 commit into
Draft
research(nightly): multi-subspace HNSW with coherence-weighted fusion#556ruvnet wants to merge 1 commit into
ruvnet wants to merge 1 commit into
Conversation
…usion Introduces `ruvector-subspace-hnsw` (ADR-199): K independent HNSW indexes on D/K-dimensional subspaces fused via per-query variance-based coherence weights. Measured results (release, x86-64, N=10K D=128): Baseline-HNSW: recall@10=0.543, 184µs mean, 5422 QPS, 6.59MB SubspaceUnion: recall@10=0.443, 874µs mean, 1144 QPS, 16.53MB CoherenceHnsw: recall@10=0.443, 880µs mean, 1136 QPS, 16.53MB Unit-test scale (N=2K D=64): coherence fusion +21pp vs baseline (0.840 vs 0.630). - 15 unit tests pass (cargo test -p ruvector-subspace-hnsw) - Benchmark binary passes acceptance thresholds - ADR-199 documents scale-dependent coherence benefit and production path - Research doc covers SOTA (Subspace Collision arXiv:2411.14754, TaCo arXiv:2603.24919) - Public gist at docs/research/nightly/2026-06-12-multi-subspace-hnsw/gist.md https://claude.ai/code/session_014erBW46sqeMFV8ipF4XsBy
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Adds nightly RuVector research for multi-subspace HNSW with coherence-weighted fusion.
What this adds
Introduces
crates/ruvector-subspace-hnsw— a new standalone crate implementing three measurable ANN variants:w_s = 1/(1+CV_s))Key research finding
Coherence fusion provides a +21pp recall improvement at N=2K, D=64 compared to single-space HNSW (0.840 vs 0.630). At N=10K D=128, the single-space baseline dominates — the scale-dependent crossover is a concrete open research question.
Real benchmark results (N=10K, D=128, x86-64)
All numbers from
cargo run --release -p ruvector-subspace-hnsw --bin benchmark. No fake tables.Includes
crates/ruvector-subspace-hnsw/) — pure Rust, onlyrand = "0.8"depcargo test -p ruvector-subspace-hnsw)docs/adr/ADR-199-multi-subspace-hnsw-coherence-fusion.md)docs/research/nightly/2026-06-12-multi-subspace-hnsw/README.md)docs/research/nightly/2026-06-12-multi-subspace-hnsw/gist.md)Prior art gap
Closest prior work: Subspace Collision (arXiv:2411.14754, SIGMOD 2025) and TaCo (arXiv:2603.24919, March 2026) use clustering indexes with static collision-count fusion. This is the first implementation using HNSW per subspace with runtime variance-based coherence weighting.
Research doc:
docs/research/nightly/2026-06-12-multi-subspace-hnsw/README.mdADR:
docs/adr/ADR-199-multi-subspace-hnsw-coherence-fusion.mdGenerated by Claude Code