AiSAQ

title: AiSAQ slug: aisaq category: Research Papers & Surveys tags:

• ann
• similarity-search
• vector-indexing source_url: https://arxiv.org/pdf/2404.06004.pdf featured: false

Overview

AiSAQ (All-in-Storage ANNS with Product Quantization) is a research method for approximate nearest neighbor search (ANNS) that places compressed vectors entirely on SSD, enabling DRAM-free (or near-DRAM-free) vector similarity search at billion-scale. It builds on DiskANN, modifying how product-quantized vectors are stored and accessed to drastically cut RAM usage while maintaining high recall and practical latency.

Key Details

• Type: Research paper / algorithmic method
• Domain: Large-scale vector search, information retrieval, RAG backends
• Core idea: Offload PQ-compressed vectors from DRAM into SSD-based indices, so memory usage no longer scales with dataset size.
• Code: DiskANN-based implementation available on GitHub: https://github.com/KioxiaAmerica/aisaq-diskann

Features

• All-in-storage PQ design

  • Compressed (product-quantized) node vectors are stored on SSD instead of being kept in DRAM.
  • Breaks the proportional relationship between DRAM usage and dataset size.

• Extremely low DRAM footprint

  • Achieves on the order of ~10 MB memory usage for query search on billion-scale vector datasets (as reported in the paper abstract).
  • Suitable for environments where DRAM is costly or limited.

• Based on DiskANN

  • Uses DiskANN as the underlying graph-based ANNS framework.
  • Preserves the graph-search paradigm and re-ranking strategy, but changes where/how compressed vectors are stored.

• Product Quantization (PQ) for compression

  • Employs PQ to represent high-dimensional vectors compactly.
  • Relieves the DiskANN trade-off where increasing compression lowers both memory use and recall, by moving PQ data to storage.

• Maintains recall–latency balance

  • Designed to achieve DRAM-free or near-DRAM-free search “without critical latency degradation” compared to standard DiskANN setups.
  • Still uses full-precision vectors from storage for re-ranking along the search path.

• Fast index switching

  • Reduces index load time required before queries can be served.
  • Makes it practical to switch rapidly between multiple billion-scale indices, useful when many vector collections must be queried selectively.

• Suitable for RAG (Retrieval-Augmented Generation)

  • Can act as a retriever backend for LLM-based RAG systems.
  • Multiple external knowledge sources can be stored as separate indices and switched on demand, without loading all index data into RAM.

• Scalability and multi-server deployment

  • Intended to scale out across multiple-server systems for emerging, very large datasets.
  • SSD-based index design fits distributed or sharded deployments.

• Use with vector database systems

  • Conceptually related to existing DiskANN-based services used in vector databases such as Weaviate and Zilliz (mentioned as context in the paper).

Use Cases

• Large-scale image, music, or document retrieval where dataset size reaches billions of vectors.
• RAG systems that:
  • Need to query multiple knowledge bases / indices.
  • Require fast index switching without reloading large indices into DRAM.
• Cost-sensitive deployments where minimizing DRAM footprint is essential while still needing high-quality ANN search.

Pricing

• This is a research method / open implementation, not a commercial product.
• No pricing or commercial plans are specified in the paper content.