research(nightly): semantic-search-cache — ANN result-set semantic cache in Rust

[ruvnet]

Summary

• Adds ruvector-semantic-cache: zero-dependency Rust crate implementing an ANN result-set semantic cache keyed by cosine similarity
• Benchmarks three variants — NoCache, SemanticCacheCoarse (t=0.90), SemanticCacheFine (t=0.97) — against a 50 K × 128-dim corpus with realistic mixed agent workload
• Real measured results: 72.8% hit rate, 92 µs hit latency (vs 7.9 ms corpus scan), 3.5× end-to-end speedup, 94.7% hit recall@10 — all 5 acceptance tests PASS
• ADR-296 records the architectural decision and production migration path
• Research doc covers 2026 SOTA (QVCache arXiv 2602.02057), 10–20 year thesis, and ecosystem integration plan

Deliverables

Artifact	Path
Crate	crates/ruvector-semantic-cache/
Research doc	docs/research/nightly/2026-08-05-semantic-search-cache/README.md
ADR	docs/adr/ADR-296-semantic-query-cache.md
Public gist	docs/research/nightly/2026-08-05-semantic-search-cache/gist.md

Benchmark results (real numbers, 2026-08-05)

Corpus: 50 000 × 128-dim f32 | Queries: 2 400 bench | k=10 | cache=500 entries
Workload: 35% exact repeats | 40% near-dup (σ=0.02) | 25% diverse (σ=0.10)

Variant                  HitRate%  MeanLat µs  p50 µs   HitRecall@10
NoCache                      0.0%    7 925.8   7 885.4        —
SemanticCacheCoarse(t=0.90) 72.8%    2 263.0      97.8    0.947  ✅
SemanticCacheFine(t=0.97)   52.3%    3 921.8     157.6    0.958  ✅

ACCEPTANCE: PASS — all 5 tests passed.

Tests

cargo test -p ruvector-semantic-cache
running 12 tests … test result: ok. 12 passed; 0 failed

SOTA basis

• QVCache (arXiv 2602.02057, Feb 2026): first paper to cache ANN result sets. No Rust implementation existed.
• vCache (arXiv 2502.03771): per-query adaptive thresholds, formal error bounds.
• Category-Aware Caching (arXiv 2510.26835): per-workload-type threshold partitioning.

Gap addressed

Cache invalidation on mutable indices (all prior systems assume static corpora) is an open problem. This PoC establishes invalidate_all() as the conservative correct baseline.

---

Generated by Claude Code