OceanBase V4.5.0 CE: AI-Native Vector Search and Kernel Enhancements

by

OceanBase V4.5.0_CE, released on December 3, 2025, positions itself as an “AI-native” iteration. The release notes highlight substantial upgrades to vector search capabilities and kernel performance, aiming to simplify AI workloads within the database layer.

AI and Vector Search Upgrades

The core focus of V4.5.0 is vector search optimization. OceanBase has enhanced IVF (Inverted File Index) and IVF_PQ (Product Quantization) indexes across four dimensions: build, query, DML, and usability.

  • Build Optimization: Index construction time is optimized through Kmeans parallelization and domain ID merge frameworks.
  • Query Optimization: An iterative filtering algorithm is introduced to intelligently select pre/post filter strategies based on different data filtering rates, balancing result completeness with speed.
  • DML Optimization: Incremental write efficiency is improved via a caching mechanism.

Additionally, the release introduces an experimental memory-based sparse vector index. Users can specify lib=vsag,type=sindi to create this index. The vendor claims this offers superior performance to disk-based sparse vector indexes when sufficient memory is available.

To lower the barrier to entry for vector search, OceanBase is introducing semantic indexing (experimental). This feature leverages built-in Embedding capabilities to make vector concepts transparent to users. Users can store raw text; the database automatically handles the conversion to vectors and indexing during writes and queries.

Hybrid search capabilities have also been expanded. The release supports multi-vector queries combined with scalar conditions, allowing for weighted fusion of result sets to refine search relevance.

Kernel and Performance Improvements

Several kernel-level optimizations have been made:

  • Index Skip Scan: This feature is now enabled by default. The vendor claims performance degradation is minimal even if the execution plan incorrectly selects this path, and the optimizer will automatically generate these plans when statistics are collected and conditions are met.
  • JSON Multi-Value Index: This index type now integrates with the Index Merge framework. By merging scan results from multiple indexes before table lookups, the system significantly reduces I/O for filtering operations involving member of, json_overlaps, and json_contains.
  • Multi Get & Insert Up: Performance for retrieving multiple rows and for upsert operations has been optimized, particularly for large datasets and scenarios with good data locality.

New AI Functions

The release introduces new SQL functions to support prompt engineering and text generation:

  • AI_PROMPT: A function template that allows prompts to be parameterized and reused, accepting placeholders like {0}.
  • AI_COMPLETE: A function that calls a text generation model for tasks like sentiment analysis or summarization.
  • Similarity Expressions: New functions such as inner_product_similarity, cosine_similarity, and l2_similarity allow direct calculation of vector distances.

Compatibility and Upgrade Path

The release includes several compatibility changes:

  • Compression: The compression function zstd_1.5.7 is added to several configuration options.
  • Storage Format: The default micro block format version is bumped from 1 to 2, enabling the new Flat format by default.
  • Views: New columns have been added to views like DBA_TABLES and DBA_OB_USERS to track replica types and plugin information, respectively.
  • System Variable: plsql_can_transform_sql_to_assign now defaults to True.

Upgrading from V4.4.1_CE to V4.5.0_CE is supported via a smooth upgrade path.