MongoDB launches latest Voyage models to aid AI development

Modeling for success

Most AI initiatives still fail to make it past the pilot stage despite the emphasis many enterprises have placed on AI development over the past few years, coupled with attempts by data management and AI vendors to create environments within their platforms that make it easy to build AI tools.

This matters because it attacks the biggest practical failure point in AI initiatives, which is the gap between promising demos and systems that actually run the business.

Fragmented systems and disorganized data estates that make relevant data difficult to discover are among the myriad problems many enterprises face when trying to develop agents and other applications.

Databricks recently launched Instructed Retrieval, an alternative to retrieval-augmented generation pipelines that augments user queries with additional specifications to better enable users to discover relevant information.

MongoDB is similarly attempting to improve data retrieval. However, rather than develop a new method of retrieving data, the vendor is working to improve the models that categorize data to make it discoverable. The move was driven by conversations with customers, according to Ben Cefalo, senior vice president and head of core products at MongoDB.

"Over the past few months, we've spent time with countless customers to understand where things break as AI moves from prototype to production," he said during a virtual press conference on Jan. 12. "Those conversations start with AI models. … As AI moves from demos to production-grade applications, we saw the same pattern again and again -- retrieval was fragmented and accuracy suffered."

The Voyage 4 series of models represent MongoDB's attempt to improve retrieval to foster more successful AI development. Meanwhile, natively integrating the Voyage models is aimed at providing a unified data platform that enables AI development without forcing customers to piece together pipelines, Cefalo continued.

"In an AI world, a database is not enough," he said.

Vector search and storage quickly became an essential element of AI pipelines after OpenAI's November 2022 launch of ChatGPT sparked surging interest in generative AI (GenAI).

Vector embeddings are numerical representations of data -- including unstructured data such as text and images -- that make it easily discovered. In addition, because the numerical representations symbolize data's essential characteristics, vector embeddings enable similarity searches to improve the relevancy of search responses.

Embedding models automatically assign vector embeddings. Reranking models, meanwhile, refine and reorder lists of items such as search results to make AI outputs more relevant.

MongoDB's new Voyage 4 models each include embedding and retrieval capabilities aimed at improving retrieval accuracy in AI pipelines, but with nuances. The general-purpose voyage-4 model balances retrieval accuracy, cost and latency; the voyage-4-large model delivers the highest retrieval accuracy; voyage-4-lite is optimized for cost and latency; and voyage-4-nano is designed for local development and testing.

In addition, MongoDB launched voyage-multimodal-3.5, which enables users to assign vector embeddings to video in addition to the text and images supported by voyage-multimodal-3.

Given that the models aim to improve retrieval accuracy, they are significant for MongoDB users, according to Stephen Catanzano, an analyst at Omdia, a division of Informa TechTarget.

"The launch of the Voyage 4 series of embedding models enables users to achieve state-of-the-art retrieval accuracy while optimizing for cost and latency," he said. "These models also introduce … video processing, which expands the scope of applications developers can build without requiring extensive architectural changes."

McKnight similarly noted that beyond forming a data intelligence layer in conjunction with operational data and vector search, the Voyage 4 models are valuable because they improve retrieval accuracy and process video content in addition to text and images. In addition, they are important additions because allow users to switch models without having to rewrite code, he said.

Additional capabilities

Beyond the new Voyage models, new MongoDB capabilities include the following:

• Automatically generated vector embeddings in MongoDB Vector Search, eliminating the need for separate embedding pipelines; a similar feature is in public preview for the MongoDB Community Edition.
• An AI assistant in MongoDB Compass and Atlas Data Explorer that enables users to interact with their systems using natural language.
• Lexical Prefilters for Vector Search, which provides advanced filtering capabilities for developers building semantic search interfaces.
• A new unified web interface for Atlas Data Explorer that enables users to develop sophisticated queries with AI assistance across all MongoDB Atlas clusters.
• An AI skills certification to help data and AI teams scale their data strategies and accelerate development cycles.

Beyond the Voyage AI models, automatically generated vector embeddings and the AI-powered assistant are MongoDB's highlight new features, according to McKnight.

"Together, these capabilities drastically lower operational overhead, allowing teams to transition from prototype to mission-critical production with a much simpler architecture," he said.

Catanzano likewise named automatically generated vector embeddings and the intelligent assistant the most significant new features.

"The former eliminates the need for external embedding pipelines, simplifying architecture, while the latter provides tailored, in-app AI guidance, reducing friction for developers working on complex data operations," he said.

Regarding the overall release, Catanzano termed the new capabilities "significant" because they address critical challenges related to AI deployment. In addition, he noted that from a competitive standpoint, they help differentiate MongoDB by unifying operational data and retrieval in a single system.

"MongoDB is reducing latency and complexity compared to fragmented solutions offered by competitors," Catanzano said. "We have seen a strong push by enterprises to use data platforms with all the AI capabilities they need. MongoDB provides many of the capabilities along with a solid ecosystem of parts to fill in any gaps."

Next steps

While the new features address some of the difficulties enterprises experience when trying to build AI tools, there is more MongoDB should do to improve its competitive position, according to McKnight.

"MongoDB must improve large-scale JSON insertion operations that currently perform slower than unified platforms and enhance JSON query performance that lags behind modern unified engines," he said.

In addition, MongoDB would be wise to improve compatibility with third-party platforms to make it easier for customers to configure customized data and AI stacks, provide more guidance to users integrating MongoDB with other systems, and add integrations with developer platforms, McKnight continued.

"These improvements would directly address MongoDB's documented weaknesses in extreme-scale scenarios while maintaining its core strengths as a flexible, developer-friendly document database for general-purpose workloads," he said.

Catanzano likewise suggested that MongoDB add integrations with third-party AI frameworks and popular developer platforms.

"Additionally, focusing on real-time analytics and predictive modeling capabilities could attract new users and solidify its position as a leader in AI-driven data platforms," he said.

Eric Avidon is a senior news writer for Informa TechTarget and a journalist with more than 25 years of experience. He covers analytics and data management.