Run ONNX and TensorFlow models directly inside the database. Models run in the database. Data stays put. Queries return in milliseconds.

Native vector type with built-in embedding models, approximate nearest-neighbor indexing, and similarity search.

Combines similarity search with keyword filtering to retrieve accurate, relevant context for RAG applications.

Secure user-defined functions in Python, Scala, or JavaScript to connect your workflows directly to language models of any size.

Connect AI agents directly through the MCP Server.

SQL-based data preprocessing with built-in ML preprocessing functions and transient tables. Online feature computation runs in the database, where the data already lives.

Deploy fully on-premises on any hardware. Runs entirely on your hardware, on your network, in your environment.

Read Iceberg and Delta formats natively, with credential management and external object caching. Your data stays in open formats on infrastructure you control.

Full control over AI model endpoints and in-database inference. You choose the models, you own the deployments, and you audit the decisions.

User-defined functions in Python, Scala, and JavaScript execute inside the database security perimeter.

Encryption at rest and in transit, with dynamic data masking and column-level de-identification. Built for regulated, air-gapped, and high-compliance environments.

Processes batches of column values in a single CPU instruction. Analytical query throughput scales with hardware, not configuration.

Query execution runs against CPU cache rather than RAM. Proven to be faster than traditional in-memory databases for typical OLAP workloads.

Distributes query processing across all available cores with auto partitioning and automatic storage indexes. Queries touch only the data they need.

Uses a patented Positional Delta Tree architecture to deliver real-time updates and inserts on massive datasets without ever sacrificing the blazing-fast, low-latency read performance analytics workloads demand.

Separated storage and compute with runtime memory configuration and tiered storage support. Scale the layer that’s under pressure without touching the rest.

Ingest live data via DataConnect and query it alongside historical data with sub-second latency. One system for streaming and batch.