Databricks Inc. today is kicking off what it calls the “Week of Agents” by introducing new and enhanced products that are intended to help enterprises confidently deploy and scale artificial intelligence agents.

AI agents are autonomous or semi-autonomous programs that perceive their environment, make decisions and take actions to achieve specific goals with or without human oversight. Though most enterprises are now working with AI, many are reluctant to deploy agents in mission-critical or customer-facing scenarios out of fear their actions can’t be tracked or controlled.

Databricks said it’s responding by expanding its Mosaic AI Gateway, which helps businesses manage and govern their AI models and deployments, to work with a wide range of open-source and commercial large language models. The software is in a public preview.

The AI/BI Genie Conversation API suite, now in public preview, enables developers to embed natural language chatbots into custom-built apps or popular productivity tools like Microsoft Corp.’s Teams and Sharepoint and Salesforce Inc.’s Slack.

“People love the Genie interface quite a lot, so they asked to be able to embed it into their applications,” said Naveen Rao (pictured), Databricks’ vice president of AI. “It’s a particular kind of agent that allows you to interact with your data in a business intelligence context.”

Natural language queries

Rao said the feature is aimed at business users who need to run complex queries but aren’t proficient in Structured Query Language. The agent can be embedded in any application that can ingest an application program interface. “You can put a skin on any kind of application as long as the permissions are set up.” he said. “You can even make a JavaScript app that runs in a browser.”

The Genie requires no training because it works with metadata such as column headings that are already in tables. It understands, for example, that a column labeled “cost” probably refers to money and will include that data in a financial query.

An upgraded Agent Evaluation Review App in public preview makes it easier for domain experts to provide targeted feedback, send traces for labeling, and customize evaluation criteria.

“Evaluating customized models is a big pain point for customers,” Rao said. “Building evaluations is quite hard. You need to have enough to cover all the cases, and in enough depth to probe the system well.”

Public evaluation frameworks such as Graduate-Level Google-Proof Q&A and the Natural Language Understanding Evaluation REST API Reference are useful but don’t always provide the depth that’s needed for specific use cases, he said. “We’ve made it super-easy to define evaluations in the form of questions and answers, and also to describe the task itself,” Rao said.

Also debuting today in public preview is Provision-Less Batch Inference, a new way to run batch inference with Mosaic AI using a single SQL query. It eliminates the need for users to provision infrastructure, a feature that’s particularly useful in queries against large data corpora, Rao said.

“A query against 10 million rows requires a lot more infrastructure than one against five rows,” he said. “This automatically calculates the size of the request, spins up all the necessary compute and runs the request in batch mode to minimize costs. You don’t have to think about provisioning or scaling at all.”

Today’s announcements are the first of several agent-related introductions Databricks plans to make throughout the week.

Photo: SiliconANGLE