Introducing SQL Data Insights Pro

IBM Research just shipped something that's been years in the making for the mainframe crowd: SQL Data Insights Pro, or SQL DI Pro, generally available since March 2026. The pitch is simple even if the engineering underneath isn't. Instead of exporting sensitive Db2 for z/OS data to some separate AI pipeline, you run semantic search, similarity matching, and anomaly detection right where the data already lives.

That matters more than it sounds. Big banks and insurers running Db2 for z/OS have spent decades building compliance walls around that data. Every time you move a customer record into an external analytics stack, you create a new compliance headache and a new place for something to go wrong. SQL DI Pro's whole design philosophy is to skip that step entirely, embedding AI functions directly into SQL so a claims adjuster or fraud analyst can query free-text notes and structured tables in the same breath.

The tool builds on an earlier IBM Research project, plain old SQL Data Insights, which figured out how to turn relational rows into vectors for approximate matching. SQL DI Pro pushes that further by adding transformer-based encoders for long unstructured text — think claims records, compliance documents, customer complaints — and then aligning those embeddings with the structured-data vectors in one shared latent space. In practice that means you can compare a loan applicant's written notes against thousands of others for similarity, without bolting on a separate vector database.

Two details stand out as genuinely useful rather than just marketing gloss. First, there's an incremental retraining algorithm, so embeddings update as new data streams in instead of requiring a full, expensive retrain every time the dataset shifts. Second, IBM is leaning hard on its own silicon: the Telum processor's on-chip AI acceleration, paired with a custom deep-learning compiler called zDLC, handles the heavy lifting of embedding generation and similarity scoring. IBM Research is also testing the newer Spyre accelerator, including multi-card setups, to squeeze out more throughput for these workloads.

None of this is IBM inventing semantic search — vector databases and embedding-based retrieval are old news in the broader AI world. What's new is doing it natively inside a mission-critical mainframe database that some of the largest financial institutions on earth still run their core operations on. That's a narrower, less flashy problem than building a chatbot, but for the compliance officers who lose sleep over data residency, it might be the more consequential one.