**46 attention mechanisms grounded in 7 mathematical theories -- from Flash Attention to optimal transport -- in one crate.**

cargo add ruvector-attention

Attention is the core operation in transformers, vector search, and graph neural networks, but most libraries give you one or two flavors and call it done. `ruvector-attention` ships 46 mechanisms spanning standard dot-product, sparse (Flash, linear, local-global), geometric (hyperbolic, mixed-curvature), graph (GAT, RoPE), and mixture-of-experts -- all SIMD-accelerated with quantization support. Pick the right attention for your data shape instead of forcing everything through softmax(QK^T/sqrt(d))V.

| | ruvector-attention | PyTorch `nn.MultiheadAttention` | FlashAttention (standalone) | xFormers |

|---|---|---|---|---|

| **Mechanism count** | 46 | 1 (scaled dot-product) | 1 (Flash) | ~5 |

| **Geometric attention** | Hyperbolic, spherical, mixed-curvature | No | No | No |

| **Graph attention** | Edge-featured GAT, RoPE for graphs | No | No | Limited |

| **Optimal transport** | Sliced Wasserstein, centroid OT | No | No | No |

| **Topology-gated** | Coherence-based mode switching | No | No | No |

| **Quantization** | Per-component (8-bit E, 5-bit H/S) | Via separate tools | No | Limited |

| **Language** | Rust (with WASM target) | Python/C++ | CUDA only | Python/CUDA |

| **SIMD acceleration** | Built in (4-way unrolled) | Via backend | CUDA only | Via backend |

| Feature | What It Does | Why It Matters |

|---------|-------------|----------------|

| **Flash Attention** | O(n) memory tiled computation | Process long sequences without running out of memory |

| **Mixed Curvature Fusion** | Combines Euclidean, hyperbolic, and spherical spaces in one pass | Model hierarchies, clusters, and flat data simultaneously |

| **Optimal Transport Attention** | Uses Wasserstein distance instead of dot-product similarity | Better distribution matching for retrieval and generation |

| **Topology-Gated Switching** | Automatically picks attention mode based on local coherence | Self-adapts to data characteristics without manual tuning |

| **Information Bottleneck** | Compresses attention via KL minimization | Keeps only the signal, discards noise |

| **PDE/Diffusion Attention** | Runs heat equation on a similarity graph | Smooth, noise-robust attention for irregular data |

| **Unified Diagnostics** | Health monitoring and automatic mode selection across all 7 theories | One report tells you which attention works best for your data |

**The pure-Rust vector database engine behind RuVector -- HNSW indexing, quantization, and SIMD acceleration in a single crate.**

`ruvector-core` is the foundational library that powers the entire [RuVector](https://github.com/ruvnet/ruvector) ecosystem. It gives you a production-grade vector database you can embed directly into any Rust application: insert vectors, search them in under a millisecond, filter by metadata, and compress storage up to 32x -- all without external services. If you need vector search as a library instead of a server, this is the crate.

| | ruvector-core | Typical Vector Database |

|---|---|---|

| **Deployment** | Embed as a Rust dependency -- no server, no network calls | Run a separate service, manage connections |

| **Query latency** | <0.5 ms p50 at 1M vectors with HNSW | ~1-5 ms depending on network and index |

| **Memory compression** | Scalar (4x), Product (8-32x), Binary (32x) quantization built in | Often requires paid tiers or external tools |

| **SIMD acceleration** | SimSIMD hardware-optimized distance calculations, automatic | Manual tuning or not available |

| **Search modes** | Dense vectors, sparse BM25, hybrid, MMR diversity, filtered -- all in one API | Typically dense-only; hybrid and filtering are add-ons |

| **Storage** | Zero-copy mmap with `redb` -- instant loading, no deserialization | Load time scales with dataset size |

| **Concurrency** | Lock-free indexing with parallel batch processing via Rayon | Varies; many require single-writer locks |

| **Dependencies** | Minimal -- pure Rust, compiles anywhere `rustc` runs | Often depends on C/C++ libraries (BLAS, LAPACK) |

| **Cost** | Free forever -- open source (MIT) | Per-vector or per-query pricing on managed tiers |

| Feature | What It Does | Why It Matters |

|---------|-------------|----------------|

| **HNSW Indexing** | Hierarchical Navigable Small World graphs for O(log n) approximate nearest neighbor search | Sub-millisecond queries at million-vector scale |

| **Multiple Distance Metrics** | Euclidean, Cosine, Dot Product, Manhattan | Match the metric to your embedding model without conversion |

| **Scalar Quantization** | Compress vectors to 8-bit integers (4x reduction) | Cut memory by 75% with 98% recall preserved |

| **Product Quantization** | Split vectors into subspaces with codebooks (8-32x reduction) | Store millions of vectors on a single machine |

| **Binary Quantization** | 1-bit representation (32x reduction) | Ultra-fast screening pass for massive datasets |

| **SIMD Distance** | Hardware-accelerated distance via SimSIMD | Up to 80K QPS on 8 cores without code changes |

| **Zero-Copy I/O** | Memory-mapped storage loads instantly | No deserialization step -- open a file and search immediately |

| **Hybrid Search** | Combine dense vector similarity with sparse BM25 text scoring | One query handles both semantic and keyword matching |

| **Metadata Filtering** | Apply key-value filters during search | No post-filtering needed -- results are already filtered |

| **MMR Diversification** | Maximal Marginal Relevance re-ranking | Avoid redundant results when top-K are too similar |

| **Conformal Prediction** | Uncertainty quantification on search results | Know when to trust (or distrust) a match |

| **Lock-Free Indexing** | Concurrent reads and writes without blocking | High-throughput ingestion while serving queries |

| **Batch Processing** | Parallel insert and search via Rayon | Saturate all cores for bulk operations |

- **[ruvector-gnn](../ruvector-gnn/)** - Graph Neural Network layer for learned search

- **[Main README](../../README.md)** - Complete project overview

- **[Getting Started Guide](../../docs/guide/GETTING_STARTED.md)** - Quick start tutorial

- **[Rust API Reference](../../docs/api/RUST_API.md)** - Detailed API documentation

- **[Advanced Features Guide](../../docs/guide/ADVANCED_FEATURES.md)** - Quantization, indexing, tuning

- **[Performance Tuning](../../docs/optimization/PERFORMANCE_TUNING_GUIDE.md)** - Optimization strategies

- **[API Documentation](https://docs.rs/ruvector-core)** - Full API reference on docs.rs

**MIT License** - see [LICENSE](../../LICENSE) for details.

**Part of [RuVector](https://github.com/ruvnet/ruvector) - Built by [rUv](https://ruv.io)**

[Documentation](https://docs.rs/ruvector-core) | [Crates.io](https://crates.io/crates/ruvector-core) | [GitHub](https://github.com/ruvnet/ruvector)