Databricks Feature Trails

This post introduces the FILE type, a new native column type for Databricks designed to handle multimodal data. This capability allows for the storage and querying of diverse data types, including text, images, audio, and video, within a single column. This is a significant advancement for building AI applications that require processing and analyzing various forms of data, enabling richer insights and more complex workflows.

This post introduces strategies for managing the costs associated with AI coding agents at scale. It details the technical approaches and architectural considerations for optimizing the financial impact of deploying and operating AI agents for code generation and related tasks, building upon previous work in agent performance and cost optimization.

This post introduces the concept and implementation of 'Tool Calling' for AI models. Tool calling allows AI models to interact with external tools or APIs, enabling them to perform actions beyond generating text, such as fetching real-time data, executing code, or interacting with other services. This capability significantly enhances the utility and applicability of AI models by bridging the gap between their understanding and real-world actions.

This post details the integration of Kimi K3 from Moonshot AI into Databricks through the Unity AI Gateway. It highlights how this integration enables users to leverage advanced AI models within the Databricks ecosystem, managed and controlled by the AI Spend Controls capability. This signifies an expansion of the types of AI models and services that can be accessed and governed through the gateway, enhancing the platform's utility for AI development and deployment.

This post details the integration of AI agent orchestration with Lakebase Postgres, enabling simplified management and execution of AI workflows. It outlines how leveraging Postgres within the Lakebase architecture enhances the capabilities for handling complex AI agent interactions and data management. This post specifically focuses on the critical role of R&D data residing in the Lakehouse for these agents, emphasizing the need for efficient data access and processing to improve agent performance and reduce costs.

This post introduces the implementation of a real-time fraud prevention system for government benefits. It details the architecture and technical considerations for processing transactions and identifying fraudulent activities as they occur, enabling immediate action to mitigate losses. The system likely involves stream processing, anomaly detection, and integration with existing benefit disbursement systems. This post specifically focuses on the deployment of 'agents' for production lines, enabling trusted, real-time decision-making within these systems.

The FDA has developed and deployed an AI platform that has achieved widespread daily adoption (85%) among its staff. This platform likely involves machine learning model deployment, user interface development for accessibility, and integration into existing workflows to support various FDA operations. The success metric of daily usage indicates a robust and valuable system that has been effectively adopted by its intended users.

This post introduces the concept and implementation of intent-based authorization within Omnigent, a system designed for managing and securing access. It moves beyond traditional role-based access control (RBAC) to a more granular and context-aware authorization model. The core idea is to define authorization policies based on the 'intent' of an action rather than just the identity of the user and the resource. This allows for more flexible and secure access control, especially in complex microservice environments. The post likely details the architecture of this intent-based system, including how intents are defined, evaluated, and enforced, and the technical challenges overcome in its development and deployment.

This post details the integration of Amazon S3 data access into Databricks using delegated IAM permissions. It outlines the technical implementation of allowing Databricks to securely access S3 buckets without requiring users to manage credentials directly, enhancing the self-serve infrastructure provisioning capabilities by simplifying data access for various workloads. This post extends this by detailing how Dow built a carbon footprint ledger on Databricks to accelerate sustainability at scale, leveraging the platform's capabilities for data management and analysis.