PatANN - Work in Progress #

PatANN continues development with several improvements to the core library. PatANN is expanding with additional search algorithms, deployment options, and scaling solutions. These developments focus on enhancing performance and scalability for demanding vector search applications.

If you like to stay updated, please Star the PatANN repository to receive notifications about updates and development progress.

Pattern Search Algorithm Variations #

We are currently testing several variations of core pattern search algorithm. Our initial tests show improvement in recall without latency impact but still in early stage to conclude anything for now:

• Adaptive Graph Traversal (AGT): Improved recall without latency impact
• Multi-focal Priority Queue: Better performance on high-dimensional dataset
• Dynamic Beam Width: Automatically adjusts search parameters based on query complexity

Improved Duplicate Detection and Prevention #

We are implementing a more sophisticated approach to handle duplicates when adding vectors to on-disk indices to reduce index bloat.

Client API and Hosting #

Current development focus:

• The updated cloud and thin client APIs will allow users to host PatANN on cloud in single and distributed node configurations
• We will be offering both REST and mesibo real-time APIs for efficient communication
• Users can self-host on their preferred cloud infrastructure
• We also plan to collaborate with infrastructure providers to offer hosted solutions
• Additional language SDKs (primarily based on SWIG)

Sharding API #

Horizontal scaling capabilities in development:

• Transparent query distribution
• Automated index partitioning
• Cross-shard result merging
• Dynamic cluster resizing

Next Steps #

Documentation and benchmarks for these features will be published once these features reach stability. Technical details and implementation guides will be available in the repository.