Databricks is getting ready to offer more support to enterprises building generative AI applications, with the addition of new Mosaic AI capabilities, it said Wednesday.
The new features, previewed at the data lakehouse provider’s ongoing Data + AI Summit, seek to simplify the training and management of generative AI applications. They join a suite of others built on its acquisition of large language model (LLM) and model-training software provider MosaicML for $1.3 billion a year ago.
Mosaic AI Agent Framework to accelerate gen AI app development
First up, Mosaic AI Agent Framework, now in public preview, is aimed at accelerating the development of generative AI applications underpinned by retrieval augmented generation (RAG), a technique useful when grounding foundational models in enterprise data.
While Databricks users have been able to build RAG on top of their platform for some time using alternative means, the Mosaic AI Agent Framework can be seen as the company operationalizing and productizing this facet of generative AI, analysts said.
“The AI Agent Framework makes it easier for enterprises to get up and running with RAG pipelines without having to roll their own embedding models, embedding algorithms, and standing up their own vector store,” said Bradley Shimmin, chief analyst at Omdia.
Further, the analyst said that Databricks’ move to preview the AI Agent Framework is in line with the market trend of all model makers pushing to build their proprietary embedding model capabilities.
“Early on, OpenAI’s embedding model really became the norm, but as we’re learning, embedding models matter as much if not more than generative LLMs in terms of ensuring that the contextual meaning of corporate information is most effectively presented to an LLM for a host of use cases,” Shimmin explained.
In November, rival Snowflake added similar capabilities to its Cortex offering.
Alongside the Agent AI Framework, Databricks has also added the AI Agent Evaluation tool, which is also in public preview.
The tool, according to the company, uses AI to check if outputs of a RAG-based application are high-quality and provides an interface to allow feedback from human stakeholders.
At the experimentation and research phase, the company had named this capability as the Mosaic AI Quality Lab.
Mosaic AI Gateway to help govern apps, models
Another addition, Mosaic AI Gateway, aims to help enterprises manage their LLMs and generative AI-based applications. It provides a unified interface to query, manage, and deploy any open-source or proprietary model, Databricks said, adding that this allows enterprises to switch the LLM that power their applications without needing to make complicated changes to the application code. Its guardrails also enable enterprises to track who is calling the model, set up rate limits to control spending, and filter for safety and personally identifiable information (PII).
The new capability, according to Omdia’s Shimmin, builds on the company’s ML Gateway API introduced last year.
“It’s basically meant to help customers streamline how developers write code for LLMs, creating an abstraction layer of consistency to better swap in and out various AI assets, such as models, data sources etc.” Shimmin said.
The AI Gateway interface also addresses “a crucial need” for enterprises to monitor and control model usage while ensuring compliance, according to Steven Dickens, practice lead of hybrid cloud at The Futurum Group.
The company is also introducing a set of tools for running LLMs and operating them, under the head of Mosaic AI Tools Catalog. Unlike the other additions, this is in private preview.
Will these new features keep Databricks ahead of Snowflake?
The new features and tools added to Databricks is expected to solidify its position in the market, analysts said, adding that the company still lags in a few areas, especially when compared with Snowflake.
Doug Henschen, principal analyst at Constellation Research, said, “Databricks is leading in the areas of AI and generative AI, but it still has a lot to prove on data warehousing and is behind Snowflake on data marketplace and data apps.”
“Snowflake has a strong platform and offers relatively more ease of use, has more revenue, better platform partnership opportunities and, as a result, what looks like a bigger addressable market opportunity,” Henschen explained.
Explaining further about the rivalry between Snowflake and Databricks, dbInsights’ chief analyst Tony Baer said that both companies are approaching AI from different starting points and are both seeking to get to the same place.
Databricks has always been the platform for data engineers and data scientists, and therefore has differentiated with its focus toward more technical practitioners since the days of MLFlow, the analyst explained, adding that Snowflake, on the other hand, came from the traditional data warehousing market and catered to business analysts.
“Both are aiming for each other’s constituencies. And so you can see it from the latest announcements, where Databricks has introduced tools for working with AI agents while Snowflake Cortex will in essence bury the complexities of RAG under the hood,” Baer said.
Separately, Amalgam Insights’ chief analyst Hyoun Park pointed out that Databricks had started branding most of its generative AI-related capabilities with the Mosaic brand name.
“This could be an effort by Databricks to increase the amount of revenue associated with MosaicML, in turn helping to justify the $1.3 billion acquisition,” Park said.
