Kingbase (CESI) Releases KES AI Database, Embedding AI Capabilities into the Database Kernel

As large-model applications extend from general scenarios to enterprises' more core scenarios, the integration of databases and artificial intelligence has become an important direction of industry evolution. Wang Shan, a leading figure in China's database discipline and a professor at the School of Information at Renmin University of China, said at the launch event that artificial intelligence is profoundly reshaping the database technology system, and that in the AI era, databases must not only ensure that data is stored accurately and computed correctly, but also provide a basis for large models to understand business.

From current practice, the constraints mainly come from three aspects: missing business semantics, unsuitable data formats, and gaps in permissions and auditing.

On October 30, China Electronics Technology Kingbase (Beijing) Technology Co., Ltd. (hereinafter referred to as CETC Kingbase) released the KES AI database in Beijing. The KES AI database released this time deeply embeds AI capabilities into the database kernel—data is stored in the database, semantics are maintained in the database, inference data (vectors and graphs) are generated in the database, and security is guaranteed in the database—so that enterprise data does not need to leave the database to support the efficient and secure implementation of the entire AI application chain.

In-database semantics make data understandable

Industry statistics show that more than 90% of the world's data is in private domains, of which 68% has never been analyzed or utilized. The problem is not insufficient data storage and computing capacity, but the lack of semantic descriptions for data—it is difficult for models to determine the business meaning of fields and records. Take the field "customer" as an example: it may refer to the end user, or it may refer to the legal entity that signed the contract, and the model cannot distinguish this on its own.

Under traditional architectures, business semantics are scattered across the interfaces of upper-layer applications, backend logic code, and the experience of business personnel, and are understood by people through a combination of the three. Large models do not need interfaces; they need a complete and accurate business semantic context.

The approach of the KES AI database is to build in a native semantic layer and write this "instruction manual" into the database: through query statements, each table and each field is annotated with business meaning, interrelationships, and calculation definitions; the system automatically collects basic information about the data, then artificial intelligence makes suggestions, and finally business personnel confirm and make them effective. The advantage of maintaining descriptions and data in the same database is that once the underlying structure changes, the semantics can be perceived synchronously, avoiding the common problems of structural asynchrony and semantic failure in external semantic layers.

In-database computing: making existing data inferable

Most enterprise data exists in the form of two-dimensional tables and is mainly oriented toward numerical computation. The data formats more suitable for AI are mainly vectors and graphs: vectors are used to "find things by meaning," retrieving content with similar meanings; graphs are used to "see relationships," answering who is related to whom and whether links are complete. Therefore, completing vectorization and graphification in the database is the most efficient path. Relationships, vectors, and graphs are processed collaboratively in the same database, reducing the burden of cross-database copying, data synchronization, and application integration.

For existing systems, CETC Kingbase provides a database migration solution that supports upgrading existing databases to the KES AI database with lower difficulty, lower risk, and lower cost, and fills in business semantic definitions within the database, enabling existing data to quickly enter a usable state.

The KES AI database integrates ONNX Runtime into the database, and models such as Embedding can be called through SQL to complete vector generation, index construction, and similarity retrieval within the database. Actual testing has proven that: under the conditions of 100 million 768-dimensional vectors and a 98% recall rate, CETC Kingbase's disk-based indexing algorithm reduces peak memory overhead for index construction and retrieval by 95.4% compared with the industry and reduces query latency by 80%.

Once data leaves the database, the original permission and audit system is interrupted. Enterprises usually access large models through the chain of "extracting data—vector database or RAG framework—large model." After data leaves the database, row-level and column-level permissions, field-level masking, and access audit policies all become ineffective.

The KES AI database uses the database's existing identity authentication, roles and access permissions, row-level access control, data labels, SQL firewall, and operation audit mechanisms to implement unified control over operations such as in-database semantic queries, vector retrieval, and graph access. Because semantic retrieval, vector generation, and graph construction are all accesses to data in this database, permissions can penetrate through to specific users; the entire process of storage, vectorization, graph construction, and semantic retrieval is completed within the database, and logs are retained centrally, making it possible to trace "who did what, when, and based on which data," meeting the traceability requirements of Level 3 of the Multi-Level Protection Scheme and industry regulation. Permission, masking, and audit rules are defined uniformly on the database side, modified in one place, and effective across the entire domain.

Around the above capabilities, CETC Kingbase has also launched an intelligent data development tool, covering more than 60 types of multi-source data access, including feature engineering, knowledge engineering, and knowledge graph engineering, supporting enterprises in transforming out-of-database data into usable knowledge; at the same time, through KES MCP and Agent Skills, it connects to mainstream AI development tools and agent frameworks, and provides NL2SQL agents and operations and maintenance agents.

In terms of industry-university-research cooperation, the "Beijing Key Laboratory of Intelligent Database Systems and Applications," jointly applied for by Renmin University of China and CETC Kingbase, has been approved for establishment, providing a higher-level platform for the two sides' basic research and industrial practice.