research(nightly): rairs-ivf — RAIRS IVF, ruvector's first Inverted File Index (ADR-193)

[ruvnet]

Summary

• Implements RAIRS IVF (arXiv:2601.07183, SIGMOD 2026) as crates/ruvector-rairs — ruvector's first Inverted File Index (IVF) family
• Three measurable variants behind a common AnnIndex trait: IvfFlat, RairsStrict, RairsSeil
• RairsSeil achieves 93.1% recall@10 at nprobe=1 vs IvfFlat's 61.3% — +31.8 pp — with identical memory (2,571 KB)
• 10/10 unit tests pass; cargo build --release -p ruvector-rairs green
• Gist: https://gist.github.com/ruvector/259532106aa52b14a16347075af6b6a6

Benchmark results (x86-64, rustc 1.87.0 --release, N=5K D=128)

Variant	nprobe=1 recall@10	nprobe=4 recall@10	Memory
IvfFlat (baseline)	61.3% / 26,984 QPS	97.9% / 13,532 QPS	2,571 KB
RairsStrict (SRAIR)	83.8% / 13,243 QPS	99.4% / 7,584 QPS	5,110 KB
RairsSeil (RAIRS+SEIL)	93.1% / 13,582 QPS	99.9% / 7,798 QPS	2,571 KB

Files

• crates/ruvector-rairs/ — new standalone crate (IvfFlat + RairsStrict + RairsSeil)
• docs/adr/ADR-193-rairs-ivf.md — architecture decision record
• docs/research/nightly/2026-05-12-rairs-ivf/README.md — full SOTA survey + results + roadmap

RAIRS algorithm

Each vector is assigned to a primary centroid (nearest by L2) and a secondary centroid chosen by the RAIR metric:

score(c_j) = ‖v − c_j‖² + λ · ⟨r_p, v − c_j⟩

where r_p = v − c_primary is the primary residual. This prefers secondaries that cover the opposite side of the Voronoi boundary, maximising recall for queries near boundaries.

SEIL (Shared-cell Enhanced IVF Lists) stores shared vectors in 32-entry blocks in the lower-indexed list and holds Ref { list, block } pointers in the secondary list — achieving dual-assignment recall with single-assignment memory.

Test plan

• cargo build --release -p ruvector-rairs — green
• cargo test -p ruvector-rairs — 10/10 pass
• cargo run --release -p ruvector-rairs --bin rairs-demo — real numbers captured
• Run on SIFT1M (1M×128) for production-scale validation (future)
• IVF-PQ follow-up (ADR-194 TBD)