mirror of
https://github.com/proffesor-for-testing/agentic-qe.git
synced 2026-09-19 08:45:47 +08:00
Dragan Spiridonov
6b249c034f
Implements 23 tasks across 4 phases per docs/implementation/ruvector-integration-plan.md: Phase 1 — Foundation (make what we have faster): - Native HNSW backend via @ruvector/router (150x search speedup, ADR-081) - Metadata filtering layer with composable AND/OR/NOT/FIELD expressions - Temporal tensor compression (Int8Array quantization, 4x memory reduction) - Deterministic golden-ratio dithering for cross-platform reproducibility - 15 feature flags (all default false) with CLI profiles Phase 2 — Intelligence (make agents smarter): - Neural TinyDancer router with REINFORCE policy gradient (ADR-082) - SONA three-loop enhancement (MicroLoRA 0.18μs/adapt via WASM) - Cross-domain transfer learning with Thompson Sampling (ADR-084) - Regret tracking with log-log regression growth classification Phase 3 — Safety (make agents trustworthy): - Sheaf-gated test validation via prime-radiant CohomologyEngine - Coherence-gated agent actions (3-filter pipeline, ADR-083) - Witness chain audit trail with SHA-256 hash linking - HNSW health monitoring with spectral analysis Phase 4 — Differentiation (make AQE unique): - Cognitive container export/import (RVF v2) - DAG attention scheduler for dependency-aware test ordering - CNN visual regression with spatial pooling embeddings - Behavior tree orchestration (Sequence/Selector/Parallel) - Reasoning QEC with majority-vote consensus - Browser QE dashboard scaffolding Also: remove obsolete v2-to-v3-migration test (source already removed), add ADR-081 through ADR-085, research docs, benchmarks, and integration test suites for all 4 phases. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2.5 KiB
2.5 KiB
RuVector Attention & FPGA Crates Research
Crates Covered
- ruvector-attention - 46 attention mechanisms across 7 mathematical theories
- ruvector-attention-node - NAPI-RS Node.js bindings
- ruvector-attention-wasm - WASM bindings with CGT Sheaf Attention
- ruvector-attention-unified-wasm - Unified 18+ mechanisms (Neural + DAG + Graph + Mamba SSM)
- ruvector-attention-cli - CLI with compute/benchmark/serve/REPL
- ruvector-attn-mincut - Min-cut gated attention (Dinic's algorithm)
- ruvector-fpga-transformer - FPGA inference with coherence gating
Key Findings
ruvector-attention (46 Mechanisms)
- 7 Mathematical Theories: Standard, Sparse (Flash/Linear), Geometric (Hyperbolic/Mixed Curvature), Graph (GAT/RoPE), MoE, Optimal Transport, PDE/Diffusion
- Optimal Transport Attention: Sliced Wasserstein with centroid clustering for O(M) transport
- Mixed Curvature Fused: Product manifold E^e x H^h x S^s with per-component quantization (8/5/5-bit)
- Topology-Gated: Adaptive 3-mode policy (Stable/Cautious/Freeze) with hysteresis
- Unified Diagnostics: Auto-recommends best attention mode based on data characteristics
ruvector-attn-mincut
- Uses Dinic's max-flow to prune low-value attention edges
- Temporal hysteresis prevents gate mask oscillation
- SHA-256 witness chain for deterministic verification
- 15-40% KV-cache reduction, 10-25% RSS reduction, <1% coherence degradation
ruvector-fpga-transformer
- 3 compute classes: Reflex (1-2 layers), Associative (4-6), Deliberative (all)
- Coherence gating with GateDecision: RanFull/EarlyExit/Skipped
- Fixed-point Q15 arithmetic for deterministic FPGA paths
- Ed25519 signed model artifacts
- 4 backends: NativeSim, WasmSim, FpgaDaemon (TCP), FpgaPCIe (mmap)
ruvector-attention-unified-wasm
- 18+ mechanisms in single WASM package: 7 Neural + 7 DAG + 3 Graph + Mamba SSM
- Mamba SSM: O(n) selective state space model with Hybrid Mamba-Attention (70/30 mix)
- DAG attention: Topological, Causal Cone, Critical Path, MinCut-Gated, Hierarchical Lorentz, Parallel Branch, Temporal BTSP
AQE Integration Opportunities
- FPGA Reflex/Associative/Deliberative maps 1:1 to AQE's 3-Tier Model Routing (ADR-026)
- CGT Sheaf Attention for hallucination detection in agent outputs
- DAG attention for optimizing test execution ordering (Critical Path, Parallel Branch)
- Min-cut gating to reduce compute costs by 15-40%
- Mamba SSM for O(n) long test sequence processing
- Native-speed attention via ruvector-attention-node for AQE's TypeScript runtime