Axoniz robot with magnifying glass searching a blueprint

Axoniz: The High-Performance Vector Database for Secure, Scalable AI

Axoniz is an advanced vector similarity search engine written in Rust, featuring quantized storage, hybrid search, snapshot recovery, and encryption. Optimized for secure and efficient AI pipelines.

Modern AI pipelines demand more from their vector databases — more speed, more scalability, and far more integrity. That’s where Axoniz comes in.

Under development for over 8 months, built from the ground up in Rust, Axoniz is a high-performance, memory-efficient, and production-ready vector similarity search engine that goes beyond standard ANN indexing. From secure quantized storage to hybrid text + vector relevance, Axoniz is designed to power the next generation of AI infrastructure.

What Makes Axoniz Different?

Axoniz is not just another vector store. It brings enterprise-grade features into a compact and modular core that supports:

  • Blazing-fast performance with Rust
  • Encrypted index storage
  • Quantized vector compression via RaBitQ
  • Hybrid BM25 + vector scoring
  • Snapshot integrity and recovery
  • Multi-tenant architecture
  • Multi-algorithm indexing
  • Flexible distance metrics

Multiple Index Types

Axoniz supports a wide range of ANN indexing algorithms for different performance, memory, and recall trade-offs:

  • Flat: Brute-force exact search with 100% recall
  • HNSW: Hierarchical Navigable Small World graph for fast approximate search
  • IVF: Inverted File index with coarse quantization
  • SPFresh: Partition-based index using the LIRE protocol for dynamic updates
  • DiskANN: Disk-based ANN optimized for massive datasets
  • RaBitQ: Proprietary 1-bit quantized index for extreme compression
    • RaBitQ HNSW: Combines quantization with graph-based retrieval
    • RaBitQ IVF: Combines quantization with large-scale IVF partitioning

Flexible Distance Metrics

  • Euclidean distance
  • Cosine similarity
  • Dot product
  • Manhattan distance
  • Hamming distance

Encrypted Indexes

All index and snapshot data are protected using:

  • AES-256-GCM
  • ChaCha20-Poly1305

With support for key rotation, versioning, and full disk-level security.

Tiered Storage

Choose your storage tier:

  • In-memory vectors for ultra-low latency
  • SSD-backed disk index with background prefetching
  • Object storage (S3, IPFS) with compression and streaming support

Snapshot System: Tamper-Resistant, Secure, and Enterprise-Grade

Security & Tamper Resistance

  • End-to-end encryption
  • Key rotation and versioning
  • Checksums and digital signatures
  • Access control enforcement

Enhanced Storage Features

  • Multipart uploads & chunked downloads
  • Adaptive compression: Deflate, Gzip, Brotli, Zstd, LZ4
  • Snapshot lifecycle management

Data Protection

  • Point-in-time consistency
  • Transactional atomicity
  • Secure deletion
  • Full audit trail

Performance Optimizations

  • Background processing
  • Parallelized I/O
  • Memory-efficient streaming

Storage Backend Support

  • Local encrypted filesystems
  • S3-compatible storage
  • NAS/network filesystems
  • Pluggable custom backend interface

Benchmark Results & Performance

Axoniz delivers exceptional performance across various workloads:

Search & Compression Efficiency

Dataset SizeDimensionsOriginal SizeCompressed SizeSearch QPSRecall@10
1M vectors3841.5 GB48 MB8,0000.97
10M vectors76830 GB960 MB5,0000.95
100M vectors1024400 GB12.8 GB3,0000.93

Operation-Level Throughput

In-Memory Operations:

  • Vector insertion: 1M+ vectors/sec
  • Vector search (HNSW): 10K+ QPS
  • Vector search (Flat): 1K+ QPS
  • Vector search (RaBitQ HNSW): 8K+ QPS
  • Vector search (RaBitQ IVF): 5K+ QPS

Disk-Based Operations:

  • Vector insertion: ~125K vectors/sec
  • Vector search: ~460 QPS

Hybrid Search (Text + Vector):

  • Combined query throughput: ~350 QPS

Memory Usage (1M vectors, 384 dimensions):

  • Full-precision: ~1.5GB
  • RaBitQ compressed: ~48MB (~32× reduction)

Comparison with Other Quantization Methods

MethodCompression RatioSearch SpeedRecall@10Training Time
RaBitQ32:1Fast0.95Medium
PQ (8-bit)4:1Medium0.92Slow
ScaNN8:1Fast0.94Slow
LSH16:1Medium0.88Fast

Use Cases

  • AI agent frameworks with persistent memory
  • Enterprise-grade RAG pipelines
  • Privacy-respecting edge AI deployments
  • Multi-tenant SaaS platforms
  • Zero-trust environments needing full encryption and rollback

Features Under Development

  • OpenTelemetry for observability
  • Cluster replication and horizontal sharding

Get Involved

Axoniz is currently in private alpha and available for early access upon request.

All core features have been implemented, with upcoming capabilities actively roadmapped. Currently assembling a team of collaborators and like minded individuals to help shape the future of Axoniz as it approaches its open-source release.

Whether you’re passionate about high-performance vector systems, secure AI infrastructure, or building cutting-edge tools in Rust — now is the time to get involved.

Interested in contributing or collaborating?
Contact me directly to become an active collaborator and contributor.