research(nightly): MUVERA FDE — multi-vector late-interaction search for ruvector#445
Draft
research(nightly): MUVERA FDE — multi-vector late-interaction search for ruvector#445
Conversation
…interaction search
Implements MUVERA Fixed Dimensional Encoding (NeurIPS 2024, arXiv:2405.19504) as a
standalone Rust crate. Converts O(n×T_q×T_d×D) brute-force ColBERT MaxSim search into
a single MIPS problem via random projection bucketing.
## Deliverables
- crates/ruvector-multivec/: new standalone Rust crate
- src/scoring.rs: maxsim_exact, chamfer_score, centroid_dot, FdeEncoder
- src/index.rs: MultiVecIndex trait + CentroidIndex, MaxSimIndex, MuveraFdeIndex,
MuveraFdeRerankIndex (FDE + exact MaxSim rerank pipeline)
- src/error.rs: MultivecError
- src/main.rs: benchmark binary (multivec-demo)
- benches/multivec_bench.rs: Criterion scoring + search benchmarks
- docs/adr/ADR-193-multi-vector-maxsim.md: architecture decision record
- docs/research/nightly/2026-05-08-multi-vector-maxsim/README.md: research document
## Key measured numbers (x86-64 Linux, cargo --release)
n=5K, T=16, D=128: FDE 38 QPS vs MaxSim 12 QPS (+3.2× speedup)
n=10K, T=32, D=128: FDE 19 QPS vs MaxSim 2 QPS (+9.5× speedup)
n=20K, T=32, D=128: FDE 9 QPS vs MaxSim 1 QPS (+9× speedup)
Criterion per-pair: centroid_dot=396ns, maxsim_exact=3.36µs, FDE_encode+dot=9.07µs (D=64)
## Tests
cargo test -p ruvector-multivec — 12 tests, all pass
https://claude.ai/code/session_018926WBoRBHzmxmQ1Qd4xSw
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Nightly research branch implementing MUVERA Fixed Dimensional Encoding (NeurIPS 2024,
arXiv:2405.19504) as a new standalone Rust crate
crates/ruvector-multivec. This closes the multi-vector search capability gap: ruvectorhad a correct but O(n×T_q×T_d×D) brute-force
MultiVectorIndex; MUVERA converts it intoa single MIPS problem, enabling standard HNSW to power ColBERT-style late-interaction retrieval.
Deliverables
crates/ruvector-multivec/— new Rust crate with 4 index variants:CentroidIndex— mean-pool baseline (cheapest, ~23% recall@10)MaxSimIndex— exact ColBERT MaxSim oracleMuveraFdeIndex— MUVERA FDE approximation (3-9× faster than brute-force)MuveraFdeRerankIndex— two-stage pipeline: FDE retrieval + exact MaxSim rerankdocs/adr/ADR-193-multi-vector-maxsim.md— architecture decision recorddocs/research/nightly/2026-05-08-multi-vector-maxsim/README.md— full researchdocument with SOTA survey, implementation notes, and real benchmark methodology
Real benchmark numbers (x86-64 Linux, cargo --release)
Criterion micro-benchmarks (D=64, T=8):
centroid_dot=396ns · maxsim=3.36µs · FDE_encode+dot=9.07µsAll 12 unit tests pass (
cargo test -p ruvector-multivec).Build is clean (
cargo build --release -p ruvector-multivec).What this is NOT
uses M=32, R=8 → 95%+ recall. This crate establishes the correct framework.
Research document
docs/research/nightly/2026-05-08-multi-vector-maxsim/README.mdADR
docs/adr/ADR-193-multi-vector-maxsim.mdPublic Gist (SEO overview)
https://gist.github.com/ruvnet/c74fbbf13352e92a0b116736dab9293f
Related
https://claude.ai/code/session_018926WBoRBHzmxmQ1Qd4xSw