feat(vector): complete vector search engine — SIMD, TurboQuant, HNSW, MVCC, IVF, GPU#27

Merged

pilotspacex-byte merged 152 commits into

mainfrom

feat/vector-engine

Apr 1, 2026

Commits on Mar 29, 2026

Commits on Mar 30, 2026

Commits on Mar 31, 2026

fix(vector): TurboQuant codebook fix, memory optimization, 2-stage HNSW search
TinDang97
committed
feat(vector): TQ-at-insert architecture — 5.5x memory reduction
TinDang97
committed
fix(vector): restore f32 reranking — recall 0.18 → 0.99
TinDang97
committed
fix(vector): mutable uses f32 L2 brute-force, not TQ-ADC
TinDang97
committed
feat(vector): TurboQuant_prod scoring — unbiased L2 without f32
TinDang97
committed
feat(vector): M=4 multi-projection QJL — recall 83% → 90%
TinDang97
committed
perf(vector): SRHT replaces dense Gaussian QJL — 25x faster insert
TinDang97
committed
fix(vector): restore dense Gaussian QJL — recall 88% → 94.5%
TinDang97
committed
feat(vector): sub-centroid refinement, multi-shard FT.*, fast insert
TinDang97
committed
perf(vector): auto-compact mutable → HNSW on first search
TinDang97
committed
perf(vector): f32 HNSW build + LUT ADC + adaptive ef — recall 87→93%, QPS 45→1126
TinDang97
committed
perf(vector): sub-centroid LUT in HNSW beam — eliminate rerank, +23% QPS
TinDang97
committed
feat(vector): configurable EF_RUNTIME and COMPACT_THRESHOLD per index
TinDang97
committed
docs: vector engine report + user guide
TinDang97
committed
feat(vector): Light/Exact build modes — Light default, 5× less memory + faster compact
TinDang97
committed
docs: vector engine report + user guide
TinDang97
committed
fix: resolve merge conflicts with main (pub/sub branch)
TinDang97
committed
fix: CI failures + address 23 code review findings
TinDang97
committed
fix: address 10 critical code review findings
TinDang97
committed
fix: remove #[inline(always)] from target_feature functions (unstable on 1.85)
TinDang97
committed

Commits on Apr 1, 2026