Bringing the power of semantic AI to IBM Db2

Enterprise customers are increasingly faced with data sovereignty, compliance, and regulatory requirements, making AI systems integration increasingly important. To support these needs, the IBM SQL Data Insights Pro (SQL DI Pro), generally available as of March 2026, introduces semantic search, similarity discovery, anomaly detection, and unified analysis of structured and unstructured data in Db2 for z/OS. In SQL DI Pro, AI moves to Db2, thereby reducing risks in areas such as privacy, compliance, and data inconsistency that arise when customers are required to move data to other analytical frameworks. From a research perspective, SQL DI Pro represents an initial pioneering product of AI and data systems integration, opening the door to further AI systems integration research.

This technology was ideated and prototyped by the IBM Research team. It extends traditional SQL from syntax-based querying to AI-driven pattern discovery, enabling developers and data analysts to uncover insights from both structured and unstructured data that were previously difficult or impossible to access.

Why semantic query in Db2?

IBM Db2 for z/OS is a mission-critical enterprise Relational Database Management System (RDBMS) trusted by many of the world’s largest enterprise institutions. While the data stored in Db2 for z/OS is rich and extensive, extracting meaningful insights becomes increasingly challenging as data volume and complexity grow.

Use cases such as similarity detection across business entities, anomaly detection in enterprise transactions, and data-driven decision recommendations require capabilities beyond traditional SQL. These insights are difficult to achieve using conventional relational queries alone.

By embedding advanced semantic analysis directly within Db2 for z/OS, SQL DI Pro enables the following:

• Deeper and more intuitive data understanding
• Reduced need for external data processing pipelines
• Lower operational cost and complexity
• Easier compliance with data sovereignty and governance requirements

Unifying structured and unstructured data insight discovery

SQL DI Pro builds on the foundation established by SQL Data Insights (SQL DI), also developed by the IBM Research team. SQL DI introduced the concept of embedding structured relational data into dense vector representations, enabling similarity-based operations such as clustering and entity comparison directly within Db2. By transforming rows into vectors in a latent space, SQL DI allowed relational data to be analyzed beyond exact-match predicates, supporting approximate reasoning over business data entities.

SQL DI Pro significantly extends this foundation to fully capture semantic meaning in unstructured data. In enterprise workloads, a substantial portion of valuable information resides in free-form text fields, such as customer notes, transaction descriptions, claims records, and compliance documents. These fields are often too large, sparse, or linguistically complex to be effectively modeled using traditional attribute-based encoding. SQL DI Pro incorporates modern encoder models from IBM to generate high-quality embeddings for columns containing long unstructured text. This enables semantic understanding that surpasses traditional structured data analysis.

SQL DI Pro integrates unstructured column embedding with structured column embedding to form a cohesive semantic layer via the following:

• Column-level embeddings for unstructured data. Long text columns are processed using transformer-based encoders to produce dense semantic vectors. These embeddings capture context, topic similarity, and implicit relationships across textual fields.
• Structured data embeddings. Tabular attributes continue to be encoded using specialized models designed for relational data, preserving numerical relationships, categorical semantics, and inter-column dependencies.
• Alignment into a shared latent space. Both structured and unstructured embeddings are normalized and projected into a common vector space, enabling direct comparison and joint reasoning across modalities.

This unified approach allows users to perform end-to-end AI-driven discovery across diverse data types in Db2 for z/OS, while applying the most suitable models for each data modality.

The following diagram illustrates this workflow in the context of an insurance loan review scenario.

Built-in semantic functions in SQL DI Pro

SQL DI Pro introduces four built-in SQL functions to support semantic analysis. These functions can be embedded into SQL statements for unified data processing.

Incremental model training for ever-changing data

SQL DI Pro also includes an incremental retraining algorithm that eliminates the need for costly full retraining of database embeddings when new data is introduced. This approach enables embeddings to be updated as new data arrives or existing data changes, significantly reducing compute and processing overhead compared to full retraining. As a result, database embeddings remain continuously up to date, ensuring that SQL DI Pro’s advanced semantic search reflects the most current patterns in an evolving dataset.

Accelerating semantic query processing with AI hardware

SQL DI Pro accelerates both embedding generation and model inference by leveraging the on-chip AI capabilities of IBM Z Telum processors, in conjunction with the IBM Z Deep Learning Compiler (zDLC) — a compiler stack developed by IBM Research to optimize deep learning workloads for z/Architecture.

This acceleration is central to making semantic query processing practical inside Db2 for z/OS. Operations such as embedding generation, similarity scoring, clustering, and anomaly detection can be computationally intensive, especially when applied to large enterprise datasets. By optimizing these workloads for IBM Z AI acceleration, SQL DI Pro brings AI closer to mission-critical data while reducing the need to move sensitive information into external analytical environments.

The Research team is also exploring the use of IBM Z with the Spyre Accelerator to further enhance acceleration capabilities for SQL DI Pro workloads. This includes investigating Spyre in multi-card settings, where semantic processing tasks can execute in parallel across multiple accelerators to improve throughput and scalability. The team is also exploring hybrid integration of both Telum and Spyre, enabling workloads to use the most suitable acceleration resource based on latency, scale, data locality, and system utilization. This hybrid acceleration direction is expected to help SQL DI Pro maximize performance and resource efficiency while preserving IBM Z’s strengths in security, governance, availability, and digital sovereignty.