research(nightly): gnn-rerank — GNN score diffusion +10.4 pp recall@10#479
Draft
ruvnet wants to merge 6 commits into
Draft
research(nightly): gnn-rerank — GNN score diffusion +10.4 pp recall@10#479ruvnet wants to merge 6 commits into
ruvnet wants to merge 6 commits into
Conversation
SOTA scan: graph-based reranking is a validated 2025-2026 direction (GNRR +5.8pp AP, Maniscope +7pp NDCG, AQR-HNSW 2.5-3.3x QPS). No production Rust vector DB applies GNN message passing over ANN candidate subgraph topology. Introduces this as the nightly research topic and documents ten candidate alternatives with scoring formula.
Implements CandidateReranker trait with four variants: - NoisyScoreReranker: passthrough baseline (28.0% recall@10) - GnnDiffusionReranker: 1-hop score averaging (+10.4 pp → 38.4%) - GnnMincutReranker: structurally-gated diffusion (38.4%) - ExactL2Reranker: oracle upper bound (74.9%) All files < 500 lines. No unsafe code. No external service dependency.
Tests cover: graph degree bounds, self-not-in-neighbours invariant, ExactL2 correctness, error conditions (empty/k-too-large), all three reranker variants produce k results. All 14 pass: cargo test -p ruvector-gnn-rerank.
Hardware: x86-64, Intel Celeron N4020, Linux 6.18.5, rustc 1.87.0 N=5000, D=128, clusters=20, K=10, retrieval_k=80, noise_sigma=0.40 NoisyScore (baseline): 28.0% recall@10, 0.2 µs, 4.9M QPS GnnDiffusion (1-hop): 38.4% recall@10, 1006 µs, 994 QPS +10.4 pp GnnMincut (coh≥0.50): 38.4% recall@10, 999 µs, 1001 QPS +10.4 pp ExactL2 (oracle): 74.9% recall@10, 13.8 µs, 72.5K QPS Candidate coverage: 74.9%. Acceptance: PASS.
Records architecture decision: CandidateReranker trait, four variants, graph construction design, acceptance of structural coherence gating over score-ratio gating, benchmark evidence, failure modes, open questions.
Public technical article covering: - GNN score diffusion algorithm and three variants - Real benchmark table from cargo run --release - Comparison with Milvus, Qdrant, Weaviate, LanceDB, FAISS, pgvector - 10 practical + 8 exotic applications - Roadmap: now/next/later - 11 cited arXiv papers (2024-2026) - 25 SEO keywords + GitHub topic tags
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.
Nightly RuVector Research — 2026-05-21 — GNN-Enhanced Candidate Reranking
Adds nightly research for
gnn-rerank: graph neural score diffusion overapproximate ANN candidate sets.
Includes
Working Rust PoC —
crates/ruvector-gnn-rerankCandidateRerankertrait with 4 variants (NoisyScore, GnnDiffusion, GnnMincut, ExactL2)ADR —
docs/adr/ADR-194-gnn-rerank.mdResearch document —
docs/research/nightly/2026-05-21-gnn-rerank/README.mdPublic gist —
docs/research/nightly/2026-05-21-gnn-rerank/gist.mdReal benchmark numbers (x86-64, Linux 6.18.5, rustc 1.87.0, N=5K, D=128, K=10)
GNN score diffusion: +10.4 pp recall@10. Acceptance: PASS ✓
Build
Main lesson
Graph score diffusion is cheap (~1ms overhead) and measurably effective (+10.4 pp).
ExactL2 reranking is faster (14µs) when full-precision candidates are available locally.
GNN diffusion's value is highest when candidate vectors must be fetched or are stored compressed.
This branch should either become a production RuVector capability or a falsified research path with useful evidence.
Generated by Claude Code