ruvnet

Novel research agenda for extending ruvLLM, RuVector, and Cognitum

Executive summary

The biggest opportunity is not one new model trick.

It is a unified selective intelligence layer across RuVector, ruvLLM, and Cognitum.

The state of the art is moving in one direction:

ruvnet

Here’s a ready‑to‑paste CI workflow that builds a plugin WASM, pushes it as an OCI artifact, signs it, verifies Rekor, checks lineage with AgentDB, and either auto‑stagestamps or opens a human‑review PR—minimal by design so you can drop it into .github/workflows/ruflo-promote.yml and try with one plugin before scaling.

name: ruflo-promote on: push: paths: - plugins/**/* branches: [ main ] workflow_dispatch: env:

ruvnet

Agentic Validation System

A three-layer regression-protection stack for AI-built codebases that ship fast across many small fixes. Designed for projects where:

• Releases are frequent (daily alphas, not monthly majors)
• Fixes are small (one-line CLI parser swaps, dependency moves) but many in number
• Multiple agents touch overlapping code in the same release window
• A regression that ships affects every user immediately, not just one customer

The stack catches three distinct regression classes that traditional CI misses, then provides forensic tools to answer "when did this break and what changed?"

ruvnet

Jujutsu vs Git Worktrees: Key Differences

TL;DR

Git Worktrees: Multiple working directories pointing to different branches of the same repo
Jujutsu (jj): Fundamentally different VCS with operation log, first-class conflicts, and automatic change tracking

They solve different problems and aren't direct alternatives.

---

ruvnet

ruvector 2026: MUVERA FDE — High-Performance Multi-Vector Search in Rust

150-char summary: ruvector adds MUVERA Fixed Dimensional Encodings (NeurIPS 2024): compress ColBERT token sets to single vectors, search at 301× brute-force speed in pure safe Rust.

Introduction

ColBERT-style multi-vector retrieval achieves state-of-the-art recall on text search tasks — but at a steep scalability cost. Every query requires computing MaxSim across all document token embeddings: for 5 million docs with 128 tokens each, that is tens of trillions of operations per query. Existing engines like PLAID work around this with custom centroid pruning infrastructure, but none generalise cleanly to standard HNSW or DiskANN indices.

ruvector-muvera implements MUVERA Fixed Dimensional Encodings (arXiv:2405.19504, NeurIPS 2024, Google Research) in pure safe Rust. FDE compresses each multi-vector document set into a single fixed-length vector via SimHash space partitioning and Rademacher random projection, enabling a

ruvnet

Pattern fill, without training

A small Rust library that copies the style of a few example sequences and produces new ones in the same shape — without training a model.

You give it a handful of examples (Mario level slices, drum loops, snippets of structured text — any short tokens that have a pattern). It reads them once. From then on it can produce new sequences that look like they came from the same source. No GPUs. No PyTorch. No model files. Just Rust.

ruvnet

Sparse-Mario

A 2,200-line Rust example that uses a subquadratic attention kernel — built for edge LLM inference on Raspberry Pi Zero 2W — as a training-free Super Mario Bros level generator. The same kernel runs in two modes from one binary:

• Autoregressive — token-by-token retrieval LM, walks the corpus's empirical bigram statistics. Now incremental via KvCache + decode_step: 2,880× faster than the original full-forward path (25 s → 9 ms for a 14×50 grid).
• Masked discrete diffusion — bidirectional context, iterative denoising with a MaskGIT cosine schedule. SOTA on this artifact: 3.8× lower L2 distance to corpus than a 1st-order Markov bigram baseline, 6.9× lower than the autoregressive path itself.

No autograd. No learned weights. No Python in the loop. The Mario corpus is the model.

---

ruvnet

You are a scheduled nightly research agent for the ruvector project. Produce deep state-of-the-art research on practical-to-exotic applications and improvements for ruvector, deliver a new feature branch with WORKING RUST code, a detailed ADR, a research document, and publish a public GitHub gist overview.

CONSTRAINTS (absolute):

• RUST ONLY. No Python, JS, TS, or anything beyond glue shell.
• NO mocks, NO TODO stubs, NO placeholder benchmarks — real cargo-run numbers only.
• Files under 500 lines. Never commit secrets. Never save to repo root.
• Use /docs for docs, /crates or /examples for code, /docs/adr for ADRs.

STEP 1 — ORIENT

• git fetch origin && git checkout main && git pull

ruvnet

ruvector 2026: MUVERA FDE — High-Performance Rust Multi-Vector Late-Interaction Search

150-word summary: ruvector now ships MUVERA Fixed Dimensional Encoding (NeurIPS 2024) as a pure Rust crate for ColBERT-style multi-vector retrieval. FDE converts O(n×T_q×T_d×D) brute-force MaxSim into a single dot-product scan, delivering 9.5× QPS improvement over brute-force at n=10K documents. Benchmark: 19 QPS vs 2 QPS (exact MaxSim oracle), x86-64 Linux, cargo --release. Three index variants — CentroidIndex, MaxSimIndex (oracle), MuveraFdeIndex — plus a two-stage FDE+Rerank pipeline.

Introduction: The Multi-Vector Search Gap in 2026

ColBERT, ColPali, and BGE-M3 have made late-interaction retrieval the dominant paradigm for precision-critical RAG pipelines. Each document is represented as T token embeddings rather than a single vector. The MaxSim score — Σ_i max_j dot(q_i, d_j) — captures nuanced semantic overlap that single-vector cosine similarity misses entirely.

The problem: scoring one query aga

ruvnet

ruvector 2026: LoRANN — High-Performance Rust Vector Search with Per-Cluster SVD Score Approximation

30.9× QPS speedup over brute-force at 56% recall@10 on 50K vectors, 54.9× at moderate recall — pure Rust, no BLAS, no Python.

ruvector now implements LoRANN (NeurIPS 2024) — a clustering-based approximate nearest-neighbour index that replaces the expensive per-cluster exact scorer with a compact rank-r SVD factorisation, achieving massive throughput gains while remaining production-deployable on commodity hardware.

Branch: research/nightly/2026-05-08-lorann · PR: #444

---