research(nightly): muvera — MUVERA multi-vector retrieval via Fixed Dimensional Encodings (NeurIPS 2024)

[ruvnet]

Nightly Research: MUVERA — Multi-Vector Retrieval via Fixed Dimensional Encodings

Date: 2026-05-08
ADR: ADR-193
Research doc: docs/research/nightly/2026-05-08-muvera/README.md
Paper: Karpukhin et al., NeurIPS 2024, arXiv:2405.19504
Gist: https://gist.github.com/ruvnet/627c64b9986a8cf1b53385708c093481

---

What

Adds crates/ruvector-muvera — the first multi-vector / late-interaction retrieval primitive in ruvector. MUVERA converts ColBERT-style multi-vector documents (one float vector per token) into a single Fixed Dimensional Encoding (FDE) whose inner product provably approximates MaxSim. After encoding, any standard single-vector MIPS index (flat scan, HNSW) works directly.

Why

ruvector had zero support for late-interaction retrieval. ColBERT and its derivatives consistently outperform bi-encoders by 3–7% nDCG@10 on BEIR benchmarks, but their multi-vector representation is incompatible with ruvector's existing AnnIndex trait. MUVERA provides the missing bridge without requiring bespoke infrastructure.

Deliverables

• Working Rust crate crates/ruvector-muvera/ — cargo build --release ✓, cargo test 11/11 ✓
• ADR-193 docs/adr/ADR-193-muvera.md
• Research doc docs/research/nightly/2026-05-08-muvera/README.md

Benchmark results (Intel Xeon @ 2.10 GHz, release build, synthetic Gaussian data)

Variant	n_docs	QPS	Build (ms)	Mem (KB)
BruteForceMaxSim	10,000	3	74	160,000
FlatFDE	10,000	14	2,441	320,000
HnswFDE	10,000	131	75,306	320,625

HnswFDE: 42.4× QPS speedup over exact BruteForce MaxSim at n=10K.

FlatFDE at n=500: 9.5× speedup with 50% memory reduction.

Architecture

FdeEncoder (R×D random projection matrix)
    └── MultiVecIndex trait
            ├── BruteForceMaxSim  — exact O(n·|Q|·|D|·d)
            ├── FlatFdeIndex      — FDE + O(n·R·D) flat scan
            └── HnswFdeIndex      — FDE + greedy HNSW, M=16

Known limitations (documented in research doc)

• HNSW build is O(n²) PoC; needs hierarchical HNSW for production (tracked as future ADR)
• Recall on random Gaussian data is low by design — FDE requires semantic embedding structure
• FDE memory overhead when R ≥ T (use R < tokens_per_doc in production)

Next steps

• Binary FDE (1-bit sign encoding, 32× memory reduction)
• IDF-weighted FDE accumulation
• Hierarchical HNSW build (O(n·log n))
• Integration with ruvector-acorn for predicate-filtered multi-vector search