Data × AI Integrated Database Engine | GBase 8c: Multi-Modal, Multi-State Data Foundation for AI Applications

With AI applications growing explosively, a database is no longer just a "place to store data"—it needs to store vectors, perform semantic searches, scale elastically, and support high-concurrency inference. GBase 8c is built on a "multi-modal, multi-state" core design, integrating vector search, HTAP, compute-storage separation, and data branching into a single system—so the data foundation never becomes a bottleneck, from AI application development to production.

Vector Engine: Query Scalars and Vectors Together with a Single SQL Statement

Almost every AI application depends on vector search. In reality, most AI applications require hybrid queries that combine scalar filtering with vector similarity. The traditional approach uses a relational database for scalars and a vector database for vectors, forcing application code to query them separately and merge results—two systems, two APIs, increasing complexity for both development and operations.

GBase 8c embeds vector capabilities directly into the database kernel. Vectors are native data types, usable in SQL just like integers and strings. Queries first filter by scalar conditions to create a bitmap, then run ANN search algorithms to find matching vector data, returning a result only when both conditions are satisfied. SQL joins and vector searches can be combined into complex, fused SQL statements, supporting window functions, stored procedures, aggregations—all standard SQL features.

One system, one SQL statement, one query—hybrid scalar and vector search done. RAG, recommendation systems, and similarity search run directly on GBase 8c, with no need for a separate vector database. It handles billions of vectors, millions of QPS, and millisecond-level query latency, all in a single system.

HTAP Row-Column Unified Storage: One System for Both Transactions and Analytics

AI applications often need real-time data—real-time recommendations require the latest transaction data, while risk models need historical behavioral data. Traditional architectures use an OLTP database for transactions and an OLAP database for analytics, with ETL moving data between them, leading to high latency and heavy operational overhead.

GBase 8c supports both row-store and column-store engines. Row-store handles transactions; column-store handles analytics. Delta tables synchronize automatically, and the query optimizer chooses the optimal path. One system simultaneously handles both TP and AP, with no data migration and near-zero latency, providing fresh data for real-time AI inference.

Unified Multi-Model Storage: Eliminate Data Silos

Data types for AI applications are becoming increasingly diverse—structured, vector, graph, time-series. Traditional approaches require a separate database for each data type, with data constantly moving between them, making consistency difficult to guarantee.

GBase 8c unifies multiple data models under a single storage management layer. One copy of data with multiple access methods eliminates data silos and reduces the complexity of developing and operating AI applications.

Compute-Storage Separation + Serverless: Handling the "Tidal" Fluctuations of AI Workloads

AI workloads fluctuate dramatically—compute power is maxed out during training, relatively stable during inference, and completely idle at other times.

The compute layer of GBase 8c is completely stateless; nodes do not persist any data, with state streamed synchronously to the storage layer via WAL. Compute can be spun up in seconds when needed and destroyed when idle. Scale down to zero when idle, driving compute costs close to zero; spin up in seconds for new workloads, with no response delay. The storage layer scales independently, supporting petabyte-level capacity with multi-copy replication ensuring 99.999% data availability.

Data Branching: A "Sandbox" Environment for AI Experimentation

AI development involves extensive experimentation—different parameters and models need to be compared and tested repeatedly. The traditional approach clones an entire environment, which can take hours or even days when data volumes are large.

Based on a WAL-based Copy-on-Write mechanism, creating a branch in GBase 8c only records a log position without copying data, completing in seconds. The new branch shares historical pages with the main dataset and copies only the modified portions upon write. Branches are completely isolated, allowing multiple experiments to run in parallel. Once validated, branches can be destroyed immediately without contaminating the main dataset.

Multi-Compatibility: Smooth Migration, Lower Migration Costs for AI Applications

GBase 8c supports multiple syntax compatibility modes, including Oracle, MySQL, PostgreSQL, and SQL Server, enabling AI applications built on different databases to migrate smoothly through a single database system. Accompanying automated migration tools support full migration, incremental synchronization, and reverse synchronization, significantly reducing migration cost and risk.

GBase 8c’s "multi-modal, multi-state" design integrates vector search, HTAP, compute-storage separation, and data branching into a single system, ensuring that the database is no longer a bottleneck for innovation in the AI era. Three real-world case studies have validated the feasibility of this approach. Looking ahead, GBase 8c will continue to deepen native AI capabilities, building a trusted data foundation for intelligent applications.