Oracle Database 23ai includes artificial intelligence (AI) vector search capabilities designed to efficiently query data based on semantic similarities. AI vector search enables applications like voice assistants, chatbots, language translators, recommendation systems, and anomaly detection systems. AI vector search also facilitates sophisticated search and recognition capabilities on images and videos.

In conjunction with Oracle Database 23ai, Oracle Exadata System Software includes AI Smart Scan, a collection of Exadata-specific optimizations capable of improving the performance of various AI vector query operations by orders of magnitude.

AI Smart Scan automatically accelerates Oracle Database AI Vector Search with optimizations that deliver low-latency parallelized scans across massive volumes of vector data. It processes vector data at memory speed, leveraging ultra-fast Exadata RDMA Memory (XRMEM) and Exadata Smart Flash Cache in the Exadata storage servers. AI Smart Scan also performs vector distance computations and top-K filtering at the data source, avoiding unnecessary network data transfer and database server processing.

In Oracle Exadata System Software release 25.1.0, AI Smart Scan includes the following enhancements:

• Top-K filtering performance and efficiency improve by maintaining a running top-K set in each storage server.

  Before Oracle Exadata System Software release 25.1.0, AI Smart Scan returns the top-K results to the database server for each region of data being processed.

  With Oracle Exadata System Software release 25.1.0, each storage server maintains a running set for each AI Smart Scan top-K operation. As the operation proceeds, each storage server only returns results that improve the previously returned results. Results outside the running top-K are filtered out and not sent to the database server. This method sends much less data to the database server for processing, reducing traffic on the storage network and improving top-K search performance.

• AI Smart Scan supports vectors with INT8 or BINARY dimensions.

  Before Oracle Exadata System Software release 25.1.0, AI Smart Scan supported vectors with high-precision (FLOAT32 or FLOAT64) dimensions. Vectors with high-precision dimensions facilitate very accurate vector comparison and search operations, requiring more memory and processing power.

  Oracle Exadata System Software release 25.1.0 extends AI Smart Scan to support vectors with INT8 or BINARY dimensions. Vectors using INT8 or BINARY dimensions trade away a modest degree of accuracy for significant efficiency and performance gains.

  For example, compared with FLOAT32, vectors with BINARY dimensions are 32 times smaller, and distance computation is up to 40 times faster. But the affect on search accuracy is minimal with some evaluations using the Massive Text Embedding Benchmark (MTEB) showing only a 6% difference in search quality. Also, vectors with INT8 dimensions are one-quarter the size of FLOAT32 (4x compression), and some MTEB evaluations show a negligible difference in search quality between them.

• AI Smart Scan projects vector distances.

  Before Oracle Exadata System Software release 25.1.0, AI Smart Scan performed filtering operations based on vector distance calculations to avoid transmitting unnecessary vector data to the database server. However, when vector distance was selected (projected) as part of the query, that vector distance calculation was performed on the database server.

  With Oracle Exadata System Software release 25.1.0, for projected vector distances, AI Smart Scan performs CPU-intensive vector distance calculations inside the storage server and sends the calculated values back to the database server. This optimization avoids sending large amounts of vector data to the database server for vector distance calculations, reducing traffic on the storage network and improving performance for queries with vector distance projections.