KIOXIA delivers a significant update to its open-source AiSAQ software, featuring advanced usability and flexibility in AI database searches within RAG systems.
KIOXIA has announced a significant update to its AiSAQ (All-in-Storage ANNS with Product Quantization) software, further advancing the usability and flexibility of AI vector database searches within Retrieval-Augmented Generation (RAG) systems. This latest open-source release introduces new configuration options that enable system architects to tune the balance between search performance and vector capacity precisely. These are two factors that are inherently at odds due to the fixed storage capacity of SSDs in modern AI infrastructure.
Flexible Performance Tuning for RAG Systems
At the core of this AiSAQ update is the ability for administrators to define the optimal trade-off between the number of vectors stored and the achievable search performance (queries per second). In practical terms, when SSD capacity is fixed, increasing search performance requires allocating more storage per vector, which reduces the total number of vectors that can be indexed. Conversely, maximizing the number of vectors means each vector uses less SSD capacity, which can impact search speed. The new controls in AiSAQ empower architects to fine-tune this balance for their specific workload requirements, all without requiring hardware changes or upgrades.
This flexibility is particularly valuable for RAG systems, where the nature of the workload can shift rapidly between high-throughput, low-latency search and large-scale vector storage. The update also broadens AiSAQ’s applicability, making it a compelling solution for other vector-intensive applications such as offline semantic search, where similar trade-offs between performance and capacity are critical.
Removing DRAM Bottlenecks
Initially introduced in January, KIOXIA AiSAQ leverages a new approximate nearest neighbor search (ANNS) algorithm that is specifically optimized for SSDs. Unlike traditional vector search solutions that rely heavily on DRAM to store index data, AiSAQ enables direct vector search operations on SSDs, dramatically reducing host memory requirements. This architectural shift allows vector databases to scale well beyond the limitations imposed by DRAM capacity, making large-scale generative AI and RAG deployments more practical and cost-effective.
As the demand for scalable AI services continues to grow, SSDs are emerging as a practical alternative to DRAM for delivering the high throughput and low latency required by modern AI systems. KIOXIA’s AiSAQ software is designed to capitalize on this trend, enabling efficient, large-scale vector search without the memory constraints that have traditionally limited system architects.
Open Source for Broader Adoption
By open-sourcing AiSAQ, KIOXIA is actively promoting SSD-centric architectures and encouraging the broader adoption of scalable AI systems. The company’s commitment to open-source technology not only increases accessibility for developers and system architects but also fosters innovation across the AI ecosystem.
Neville Ichhaporia, Senior Vice President and General Manager of KIOXIA’s SSD business unit, emphasized that the latest AiSAQ release provides advanced tools for improving both performance and capacity in scalable RAG systems. He noted that by supporting open-source development, KIOXIA is helping to drive accessibility and innovation in AI infrastructure.
KIOXIA’s latest AiSAQ update represents a step forward for organizations building scalable, SSD-based AI infrastructure. With flexible configuration options, DRAM-free architecture, and open-source availability, AiSAQ is well-positioned to support the evolving needs of RAG systems and other vector-driven AI applications.
Download and Availability
The updated KIOXIA AiSAQ open-source software is available now and can be downloaded directly from GitHub:
https://github.com/kioxia-jp/aisaq-diskann
Engage with StorageReview
Newsletter | YouTube | Podcast iTunes/Spotify | Instagram | Twitter | TikTok | RSS Feed
