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>
2.4 KiB
2.4 KiB
RuVector CNN, Sparse Inference & Domain Expansion Research
Crates Covered
- ruvector-cnn - Pure-Rust MobileNet-V3 feature extraction with SIMD/INT8
- ruvector-cnn-wasm - Browser-side image embedding
- ruvector-sparse-inference - PowerInfer-style sparse inference (2x-52x speedups)
- ruvector-sparse-inference-wasm - Browser GGUF model inference
- ruvector-domain-expansion - Cross-domain transfer learning engine
- ruvector-domain-expansion-wasm - Browser transfer learning
Key Findings
ruvector-cnn
- MobileNet-V3 Small (576d) and Large (960d) embeddings
- SIMD backends: AVX2, NEON, WASM SIMD128, scalar
- INT8 quantization kernels per architecture
- Winograd fast convolutions
- Contrastive learning: InfoNCE (SimCLR) + Triplet loss
- Pure Rust, no Python/GPU/OpenCV dependency
ruvector-sparse-inference
- Only computes active neurons (~10% handle ~90% of activations)
- Low-rank predictor (PQ factorization) in O(rd) time
- Pi Integration System: Pi-derived scale factors for avoiding quantization artifacts
- Precision Lanes: 3-bit (reflex), 5-bit (streaming), 7-bit (reasoning), Float (training)
- Graduation policies: signals move UP lanes on novelty, DOWN on stability
- W2 transpose: 83% latency reduction via contiguous memory access
- Performance: 10% active = 52x faster, 30% = 18x, 50% = 10x
ruvector-domain-expansion
- Meta Thompson Sampling with Beta priors across context buckets
- Cross-domain transfer: extract posteriors -> sqrt-dampening -> seed target -> verify
- Transfer verification gate: target improved AND source not regressed
- Population-based policy search: 8 PolicyKernel variants with elite selection + crossover
- Curiosity-driven exploration via UCB bonus
- Regret tracking: sublinear growth = learning, linear = stagnation
- Pareto front tracking for multi-objective optimization
- 3 built-in domains: Rust Synthesis, Structured Planning, Tool Orchestration
AQE Integration Opportunities
- Visual testing: CNN embeddings for screenshot comparison without pixel matching
- Sparse inference: 2-52x faster local embedding/LLM operations
- Precision lanes: Map 3/5/7-bit to AQE's Tier 1/2/3 model routing
- Cross-domain transfer: Train on one test domain (API), transfer to another (UI)
- Population search: Evolve QE agent configurations through PolicyKernel system
- Regret tracking: Measure whether AQE agents are actually learning over time