research(nightly): graph-topology-aware-quant — hub-aware SQ4/SQ8 mixed precision

[ruvnet]

Summary

Adds nightly RuVector research for Graph-Topology-Aware Quantization (TAQ): assigning SQ8 (8-bit) precision to hub nodes and SQ4 (4-bit) precision to leaf nodes in a k-nearest-neighbor graph, achieving better recall than uniform SQ4 at lower memory than uniform SQ8.

• Recall@10: 0.9652 vs UniformSQ4's 0.8194 (+14.6pp) at only +9% more memory
• Memory: 21% of f32 baseline vs SQ8's 25% — TAQ saves 16% vs SQ8 while losing only 2.1pp recall
• Acceptance test: PASS — TAQ recall ≥ SQ4 recall AND TAQ memory ≤ SQ8 memory

What's Included

1. Working Rust PoC — crates/ruvector-taq/ with 4 index variants and 15 passing unit tests
2. ADR — docs/adr/ADR-272-graph-topology-aware-quant.md
3. Research document — docs/research/nightly/2026-07-12-graph-topology-aware-quant/README.md
4. SEO gist — docs/research/nightly/2026-07-12-graph-topology-aware-quant/gist.md
5. Real benchmark results — captured from cargo run --release -p ruvector-taq --bin benchmark

Key Benchmark Numbers

Variant	Recall@10	Memory	Mem%
FullPrecision-f32	1.0000	1,250 KB	100%
UniformSQ8	0.9864	312 KB	25%
UniformSQ4	0.8194	156 KB	12.5%
TAQ-mixed	0.9652	262 KB	21%

Hardware: x86_64 Linux. Rust 1.94.1. N=5,000, D=64, Q=500, K=10.

Algorithm

1. Build k-NN graph (k=8) over N input vectors — O(N²·D) brute-force.
2. Compute in-degree per node (how many nodes list it as a nearest neighbor).
3. Classify: in_degree > threshold → hub (SQ8), else → leaf (SQ4).
4. Fit separate quantization params per class; encode and store.
5. Search: dequantize each stored vector at query time; compute squared Euclidean; return top-K.

Ecosystem Connections

• ruvector-graph + ruvector-mincut — graph primitives for topology discovery
• ruvector-agent-memory — TAQ as compression backend for episodic memory
• cognitum-gate-kernel — edge deployment with TAQ as memory substrate
• ruvector-proof-gate — combine for precision-upgrade access control
• rvf / RVF format — serialize TAQ index to portable cognitive package
• MCP tools — memory_compress(), memory_search_compressed() as agent primitives
• ruFlo — automated memory hygiene trigger on memory growth

Research doc

docs/research/nightly/2026-07-12-graph-topology-aware-quant/README.md

ADR

docs/adr/ADR-272-graph-topology-aware-quant.md

Build and Test

cargo build --release -p ruvector-taq
cargo test -p ruvector-taq        # 15 tests pass
cargo run --release -p ruvector-taq --bin benchmark

---

Generated by Claude Code