Skip to content

research(nightly): rvf-index-shard — portable subgraph extraction for edge AI and agent memory#544

Draft
ruvnet wants to merge 5 commits into
mainfrom
research/nightly/2026-06-06-rvf-index-shard
Draft

research(nightly): rvf-index-shard — portable subgraph extraction for edge AI and agent memory#544
ruvnet wants to merge 5 commits into
mainfrom
research/nightly/2026-06-06-rvf-index-shard

Conversation

@ruvnet

@ruvnet ruvnet commented Jun 6, 2026

Copy link
Copy Markdown
Owner

Nightly RuVector Research: RVF Index Shard

Adds nightly RuVector research for portable subgraph extraction from proximity graphs, enabling edge deployment and agent memory portability.

What this PR includes

  1. Working Rust PoCcrates/ruvector-shard: three extraction strategies, binary wire format, brute-force shard search, recall measurement. Zero external service dependencies.
  2. ADR-196docs/adr/ADR-196-rvf-index-shard.md
  3. Research documentdocs/research/nightly/2026-06-06-rvf-index-shard/README.md
  4. SEO gistdocs/research/nightly/2026-06-06-rvf-index-shard/gist.md
  5. Real benchmark results — all numbers from cargo run --release -p ruvector-shard --bin benchmark

Benchmark headline numbers (x86_64 Linux, n=1024, dim=128, 12.5% shard)

Variant Speedup Random Recall@10 Biased Recall@10 Wire KB
BFS Shard 8.1× 13.9% 79.3% 67.0
Coherence Shard 8.1× 12.5% 49.0% 66.9
Hub Shard 8.3× 11.8% 18.5% 66.4

17/17 acceptance tests PASSED.

Key SOTA references

  • "Unleashing Graph Partitioning for Large-Scale ANN Search" (VLDB 2025, arXiv:2403.01797)
  • "Down with the Hierarchy: The 'H' in HNSW Stands for 'Hubs'" (ICML 2025, arXiv:2412.01940)
  • "Portable Agent Memory" (Microsoft, arXiv:2605.11032, May 2026)
  • "LEANN: Low-Storage Vector Index for Personal Devices" (ICML 2025, arXiv:2506.08276)

Ecosystem connections

  • RVF format: wire bytes follow magic + version + typed payload pattern compatible with rvf-manifest
  • HNSW/graph: ShardExtractor trait wraps any proximity graph; Phase 2 integrates ruvector-core::HnswIndex
  • Mincut/coherence: CoherenceShard uses centroid-cosine scoring validated in ruvector-coherence
  • Edge/WASM: 67KB wire fits in WASM linear memory; no_std-ready design
  • MCP tools: shard declared as MCP resource via CapabilityManifest TLV (future)
  • ruFlo: post-task hook triggers shard extraction when agent domain shifts (future)

How to verify

git checkout research/nightly/2026-06-06-rvf-index-shard
cargo build --release -p ruvector-shard
cargo test -p ruvector-shard
cargo run --release -p ruvector-shard --bin benchmark

Research doc: docs/research/nightly/2026-06-06-rvf-index-shard/README.md
ADR: docs/adr/ADR-196-rvf-index-shard.md


Generated by Claude Code

claude added 5 commits June 6, 2026 07:29
Implements three subgraph extraction strategies from k-NN proximity
graphs: BFS (graph locality), Coherence (semantic centroid), and Hub
(topological hubs). Serializes to 67KB RVSHARD binary. All 17 tests
pass. Build: cargo build --release -p ruvector-shard.

https://claude.ai/code/session_01Fu5KBvAT3YykFeKdxac1Fi
Research document covers BFS/Coherence/Hub shard extraction, SOTA
survey (VLDB 2025, ICML 2025, arXiv 2026), benchmark results, 10-20yr
thesis, and ecosystem fit. Gist included for SEO publication.

https://claude.ai/code/session_01Fu5KBvAT3YykFeKdxac1Fi
Documents the decision to introduce portable subgraph extraction,
three extraction variants, production API shape, alternatives
considered, benchmark evidence, and migration path.

https://claude.ai/code/session_01Fu5KBvAT3YykFeKdxac1Fi
BFS shard: 8.1x speedup, 79.3% recall@10 for anchor-biased queries.
Coherence shard: 8.1x speedup, 49.0% recall. Hub shard: 8.3x speedup,
18.5% recall (routing index). Wire: 67KB per 128-node shard. All
17 acceptance tests PASSED.

https://claude.ai/code/session_01Fu5KBvAT3YykFeKdxac1Fi
Adds lock entries for rand, thiserror, and serde resolved for the
new ruvector-shard crate.

https://claude.ai/code/session_01Fu5KBvAT3YykFeKdxac1Fi
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants