--- title: "Real-time AI in SQL and Expert-Level Control In Ververica Cloud and BYOC" description: "We’ve set the new baseline: governed, absolute truth, delivered at speed. Real-time AI belongs in SQL, the lakehouse in the runtime, control with the operator. This is the new baseline." lastUpdated: 2026-10-14T07:58:16.000Z source_url: html: "https://www.ververica.com/blog/real-time-ai-in-sql-and-expert-level-control-in-ververica-cloud-and-byoc" md: "https://www.ververica.com/blog/real-time-ai-in-sql-and-expert-level-control-in-ververica-cloud-and-byoc.md" --- Ververica created [Apache Flink®](https://www.ververica.com/apache-flink-vs-ververica). We wrote the standard. And for over a decade, we have been solving what others couldn't: the edge cases that break other data platforms. Now, we’re raising the bar again with new features for our [Fully-Managed Cloud](https://www.ververica.com/product/deployment) and [BYOC](https://www.ververica.com/product/deployment/byoc) deployments of [Ververica’s Unified Streaming Data Platform](https://www.ververica.com/): AI-Native SQL. A more robust lakehouse ingestion path. Operator-level control that puts power in the hands of production teams. This release sets the new baseline for data streaming platforms: governed, absolute truth, delivered at the speed of now for cloud deployments. ## SQL Becomes AI-Native Calling AI from a streaming pipeline used to mean writing UDFs, managing HTTP clients, and handling retries by hand. The updated [VERA engine](https://www.ververica.com/product/vera) introduces seven new AI functions that run natively in Flink SQL. Configure your model provider once. The pipeline does the rest.Seven functions, zero custom plumbing: - **AI_CLASSIFY** routes text into categories of your choosing with a confidence score. - **AI_SENTIMENT** scores polarity and labels input as positive, negative, or neutral. - **AI_EXTRACT** pulls structured fields from unstructured text against a JSON Schema you define. - **AI_SUMMARIZE** compresses long text to a length you specify. - **AI_EMBED** turns text into vectors for similarity and semantic retrieval. - **AI_TRANSLATE** handles mutual translation across more than 10 languages with automatic source detection. - **AI_MASK** identifies and masks sensitive fields automatically. In stream. The last one is particularly important for regulated workloads serving industries like [finance](https://www.ververica.com/banking). With **AI_MASK**, sensitive data gets redacted in motion, not after a batch runs hours later. That’s the difference between a fully compliant pipeline and a violation waiting to be audited. For years, batch and streaming engines have made you choose: approximate counts that are fast and wrong, or exact counts that are slow and expensive. Ververica refuses this trade. A native Bitmap type with cardinality functions delivers exact deduplication in stream. Unique visitors, distinct events, audience overlap. All counted, not estimated, at speed. ## Flink CDC YAML: More Sources, Smarter Routing, Safer Failure Modes - **Postgres and MongoDB sources** get full YAML support: [**Apache**](https://www.ververica.com/ecosystem-introduction/what-is-apache-paimon) [**Paimon™**](https://www.ververica.com/ecosystem-introduction/what-is-apache-paimon) is now a first-class source inside **Apache Flink Change Data Capture (CDC) YAML**, closing the loop on the [**Streamhouse**](https://www.ververica.com/ecosystem-introduction/what-is-streamhouse) pattern. - **Catalog-aware references: **YAML jobs can reference source and sink tables already registered in your catalog. No more duplicating definitions across pipeline configs and catalog entries. - **Regex-based routing: **Standard regular expressions now drive Route table logic. Table merging and database sharding stop being a custom code problem. - **Dirty data handling: **JSON parsing now isolates malformed records instead of killing the job. **Empty schema fault tolerance** lets Apache Kafka® CDC YAML create tables with empty schemas when no source data is present. Jobs start. They wait. They do not fail before they begin. These are not vanity updates. They are the changes that decide whether your pipeline survives its first production weekend. ## Connector Enhancements This release includes new connectors for **CosmosDB** and **MongoDB Catalog**. The Kafka Canal-JSON format now parses source-database index events and exposes raw key and value as metadata, and Flink removes redundant consumer groups. The Paimon Sink supports cross-partition upsert. MongoDB sinks support partial update with better ObjectId handling. Redis gets stronger cluster mode and batched writes. Elasticsearch picks up explicit doc_as_upsert and a connection timeout parameter. ## Operational Depth For Those Who Know What Difficult Looks Like Serious streaming users require per-operator control to finely tune their deployments. Ververica offers it **now**. This is where the VERA engine gets serious about giving fine-grained control to those who run streaming jobs at scale. - **Expert Mode for resource allocation: **Dynamic resource allocation now supports per-Slot-Sharing-Group configuration. For complex deployments, this is the difference between linear scaling and a wall at 60% utilization. - **Operator-level state TTL: **A single uniform TTL punishes both ends of the spectrum. Fast operators bloat state. Slow operators lose history. Now you configure TTL per operator. Logistics data keeps its memory. High-velocity operators stay lean. - **State compatibility for SQL deployments: **Resume from the latest state and VERA detects schema and topology changes automatically. The results? Less downtime. Less re-bootstrapping. Less manual surgery. - **JDK 17 **support arrives in VERA, making it available across all deployment modes. Pick the new vera-4.5-jdk17-flink-1.20 engine version when your operators, libraries, and downstream tooling are ready. ![Ververica Start Job screen in Resume Mode with Latest State selected and State Compatibility showing Full Compatible.](https://cdn.sanity.io/images/6b38iw1w/production/96df7e68b3cf48a8846fb7a892b8d903f5f3f810-1200x675.heif) ## SQL Console Until now, ad-hoc SQL queries in Ververica’s Managed Cloud and BYOC deployments meant creating a streaming job, waiting for it to start, and reading the result from the deployment logs. The new SQL Console gives ad-hoc work a dedicated environment. Run **CALL**, **DDL**, **DQL**, **DML**, and **EXPLAIN** in the same place. Create and manage catalogs and tables. Operate on Paimon tables directly. Inspect execution plans without leaving the script. ![Ververica Platform SQL Scripts editor running a CREATE CATALOG statement for a Paimon catalog on Azure Data Lake Storage.](https://cdn.sanity.io/images/6b38iw1w/production/483b9809b53ba16e14b18126bc88f845471c1183-1200x579.heif) It also closes the gap with Apache Spark™ and Delta Lake on data lake operations. For teams using SQL as their primary control surface, the operator running an incident at midnight no longer needs a deploy to ask a question. ## This is the New Baseline Real-time AI belongs in SQL. The lakehouse belongs in the runtime. Control belongs with the operator. For regulated industries, the math is simple. You need all of your data governed, masked, retained, and queryable. This release provides it all. ## Resources - [https://www.ververica.com/product/deployment/byoc](https://www.ververica.com/product/deployment/byoc) - [https://www.ververica.com/product/deployment](https://www.ververica.com/product/deployment) - [Release notes](https://docs.ververica.com/introduction/product-updates/)