Kioxia Corporation, a global provider of memory solutions, has released an updated version of its KIOXIA AiSAQ software aimed at improving vector search efficiency in AI workloads, particularly those based on retrieval-augmented generation (RAG) systems. The open-source software now includes new configuration options that allow system architects to adjust trade-offs between search performance and the number of vectors stored within fixed SSD capacity constraints.
Originally launched in January 2025, AiSAQ leverages a solid-state drive (SSD)-optimized approximate nearest neighbor search (ANNS) algorithm. Unlike conventional systems that rely heavily on DRAM, AiSAQ performs vector search directly on SSDs, reducing memory overhead and enabling the development of more scalable AI systems.
The latest update enables flexible tuning of performance metrics. For systems where SSD capacity is fixed, achieving higher query rates typically consumes more storage per vector, limiting the total number of vectors. Conversely, prioritizing higher vector counts often results in reduced performance. AiSAQ’s new controls allow for customization of this balance, making it adaptable to different workload requirements without hardware changes.
This flexibility is particularly relevant for RAG systems and other applications that handle large volumes of vector data, such as offline semantic search. By improving control over system resources, AiSAQ helps users address performance bottlenecks and better manage storage utilization.
“As AI applications scale, the ability to balance performance and capacity becomes increasingly important,” the company stated. “AiSAQ supports this with a software-centric approach that aligns with SSD-based infrastructure.”
Kioxia continues to position SSDs as a practical alternative to DRAM in AI environments, citing advantages in scalability and power efficiency. By keeping AiSAQ open-source, Kioxia aims to support broader adoption and development within the AI and data infrastructure ecosystem.
