research(nightly): muvera-fde — MUVERA Fixed Dimensional Encodings (NeurIPS 2024)

[ruvnet]

Summary

Implements MUVERA Fixed Dimensional Encodings (arXiv:2405.19504, NeurIPS 2024, Google Research) as a new standalone Rust crate ruvector-muvera.

MUVERA compresses ColBERT-style multi-token embedding sets into fixed-dimension single vectors via SimHash space partitioning + Rademacher random projection, enabling standard HNSW/IVF indexing for multi-vector workloads with a formal ε-approximation guarantee on Chamfer/MaxSim similarity.

Gist: https://gist.github.com/ruvnet/d5783fe3ec893249b1f7590f00a2c087

Key Results (x86_64, cargo --release, 4 CPUs)

Criterion micro-benchmarks (1K docs, d=128, 32 tokens/doc)

Variant	Latency	Speedup vs BruteForce
BruteForce MaxSim	61.8 ms/query	1×
FDE-small (B=8, dp=8, R=4)	205 µs/query	301×
FDE-medium (B=16, dp=16, R=4)	865 µs/query	71×
FDE-large (B=32, dp=16, R=4)	1.87 ms/query	33×

Demo (5K docs, clustered embeddings, 50 clusters σ=0.25)

Variant	Recall@10	QPS	Memory	Speedup
BruteForce	1.000	13	39.06 MB	1×
FDE-small	0.098	1,043	4.88 MB	80×
FDE-medium	0.169	257	19.53 MB	20×

Encode latency per document

Config	Latency
B=8, dp=8, R=4	49 µs/doc
B=16, dp=16, R=4	178 µs/doc
B=32, dp=16, R=4	459 µs/doc

Deliverables

• crates/ruvector-muvera/ — FdeEncoder, MuveraIndex<B: VectorBackend>, VectorBackend trait, FlatBackend, 12/12 tests passing
• docs/adr/ADR-193-muvera-fde.md — ADR with context, decision, alternatives
• docs/research/nightly/2026-05-08-muvera-fde/README.md — full research doc (SOTA survey, algorithm walkthrough, benchmark tables, roadmap)

Architecture

FDE(token_set) = concat(
  for r in 0..R:
    for b in 0..B:
      Φ_r · centroid_b(SimHash_r(token_set))
)

VectorBackend trait makes HNSW/IVF backends pluggable — see ADR-193 §Alternatives.

Tests

cargo test -p ruvector-muvera   # 12/12 pass
cargo build --release -p ruvector-muvera  # OK
cargo bench -p ruvector-muvera  # criterion green

See docs/research/nightly/2026-05-08-muvera-fde/README.md and docs/adr/ADR-193-muvera-fde.md for full analysis.

https://claude.ai/code/session_01393yTCKC5VvRYFxnZ38KH6