Snowflake Inc. is using its annual Snowflake Summit to introduce enhancements to Cortex AI, a managed service that makes it easier for organizations to discover, analyze and build artificial intelligence applications in the Snowflake Data Cloud. New features include the ability to quickly develop chatbots for querying data, a no-code interactive interface, access to several additional large language models and fine-tunings to serverless functions.

The company is also introducing a machine learning accelerator called Snowflake ML, which aims to help developers build, discover and govern models and features across the machine learning lifecycle. Developers are getting improved Python support, a new set of notebooks and native observability capabilities.

“There is a large role for Snowflake to play at every layer of the AI stack, whether it’s migration, creating data pipelines faster or constructing documentation, which none of use likes,” said Chief Executive Sridhar Ramaswamy. “Language models don’t get bored.”

Cortex Analyst and Cortex Search, both of which will soon go into public preview, enable users to build chatbots quickly. “Everything flows seamlessly, so it obeys all the rules that have been put on data,” Ramaswamy said. “Things work as expected. There are no surprises.”

New AI-powered object descriptions, which automatically generate relevant context and comments for tables and views, will enter private preview soon. Cortex Analyst, which is based on Meta Platforms Inc.’s Llama 3 and Mistral Large models, allows businesses to securely build applications on top of their analytical data in Snowflake.

Enhanced document search

Cortex Search uses technology from search startup Neeva Inc, which Snowflake acquired last year, along with the Arctic Embed suite of text embedding models focused on retrieval and performance. The combination allows users to build applications that search against documents and other text-based datasets using both vector and text and deliver the capability as a service, Snowflake said.

Users don’t need to set up, integrate or manage a separate vector store to build retrieval-augmented generation or low-latency search. The feature creates a service using a single function or no-code interface in Snowflake’s AI & ML Studio, and the chatbot can be integrated into any application using the representational state transfer or Python application program interface.

Cortex Guard leverages Meta’s Llama Guard LLM-based input-output safeguard to filter and flag harmful content to reduce generative AI hallucinations and mistakes. “It understands whether the information asked for is in the document is in the parameters of the LLM,” said Baris Gultekin, the company’s head of AI. “Knowing the data isn’t in the document, it can reject the answer.”

Pre-built AI apps

Snowflake is also rolling out a set of pre-built AI-powered applications based on its own models. The forthcoming Document AI lets users extract content like invoice amounts or contract terms from documents using Snowflake’s multimodal Arctic-TILT LLM, which is tailored to understand and extract data from documents. The company also said its Copilot, which was announced last November, will be generally available soon. It can combine Mistral Large with a proprietary SQL generation model to improve SQL queries.

The new Snowflake AI & ML Studio is a no-code interactive interface now in private preview for moving AI applications into production faster. It can be used to test and evaluate models prior to deployment quickly. Cortex Fine-Tuning, which is entering public preview, can be accessed through AI & ML Studio or as an SQL function. Serverless customization is available for a subset of Meta and Mistral AI models, which can be accessed through a Cortex AI function managed by Snowflake role-based access controls.

Enhancements to the core Snowflake platform are intended to improve flexibility and interoperability regardless of where data resides. The big announcement is that support for Apache Iceberg Tables is now generally available. Iceberg is gaining favor as a standard for data lakes because of its ability to evolve table schemas over time without rewrites, flexible partitioning and a feature called “time travel” that allows queries to be run against historical data, among other features. Unified Iceberg Tables, which were introduced last August, let customers work with their own Iceberg data managed by Snowflake.

Availability was timed to coincide with yesterday’s announcement of Polaris Catalog, a vendor-neutral cross-engine catalog for Iceberg that will be released to open-source this summer.

Internal marketplace

An internal marketplace for the Snowflake Horizon governance platform is entering a private preview. It allows users to curate and publish data products such as data, models, and applications for others in their organization to discover without exposing them externally. Sharing capabilities entering private preview soon will cover AI models, Iceberg Tables, and Dynamic Tables.

The company also said it’s expanding its AI Data Cloud to address highly regulated and sovereign markets. This includes a European Union-only data boundary that keeps all customer data and relevant service and usage data within regional borders to satisfy data residency and sovereignty requirements. The company will also offer a separate environment to Department of Defense customers, including a networking integration with the DoD’s Boundary Cloud Access Point secure gateway.

New tools for AI development

Snowflake Notebooks, introduced last November, have entered public preview with native integration with the full Snowflake platform, including the Snowpark ML execution environment, Streamlit development tools for Python and Cortex AI artificial intelligence development. Snowflake Notebooks provides a single development interface for Python, SQL and Markdown syntax. Developers can use them to iterate on machine learning pipelines with AI-powered editing features that the company said simplify data engineering workflows.

Snowpark’s new application program interface supports the pandas open source data analysis and manipulation library. Now in public preview, it is said to enable Python developers to work with pandas syntax for advanced AI and pipeline development within the Snowflake environment.

The company said it takes a data-centric approach to the DevOps methodology by integrating development, operations and data management within a single platform.

“We see a trend toward more applications trying to pull data from Snowflake, and that complicates security,” said Christian Kleinerman, executive vice president of product. “Instead of taking data to a number of applications, we want to bring the business logic to the data. It’s easier to bring over a few thousand lines of code than to copy terabytes of information.”

Snowflake is championing a declarative approach to development with a new Database Change Management feature now in public preview. Developers will soon be able to use Git integration for collaborative development and streamlined deployment across different environments. A Snowflake Python API will be available soon to manage resources and the forthcoming open-source Snowflake Command Line Interface will provide a single window to manage CI/CD pipelines.

Built-in observability

The new Snowflake Trail is a set of observability capabilities for data quality, pipelines and applications that enable developers to monitor, troubleshoot, and optimize workflows more easily. Built-in telemetry signals are provided for Snowpark and Snowpark Container Services encompassing metrics, logs, and distributed tracing without the need for manual setup. Snowflake Trail adheres to OpenTelemetry standards, which are supported by most observability platforms and integrate with many open-source visualization and alerting tools.

To further support for AI applications, the company is also announcing a public preview of integration between Snowpark Container Services and the Snowflake Native App Framework. This enables organizations to extend the breadth and variety of applications they build in the AI Data Cloud using configurable graphics processing units and CPU instances. Developers can build Snowflake Native Apps once and deploy and distribute them across clouds and regions through Snowflake Marketplace.

Photo: SiliconANGLE