Without a modern, robust data foundation, AI is a wasted investment. This sentiment is one that enterprises must reckon with as they hurdle into their AI implementations, understanding that AI cannot simply be a matter of “act now, ask later.” To succeed with AI, organizations will not only need to truly modernize their data infrastructure, but adopt a variety of best practices, technologies, and skills to drive tangible value.
DBTA’s latest webinar, Data Management for AI: Knowledge Graphs, Vector Databases, and Semantic Layers, gathered three experts to explore the nuances of AI and the importance of its underlying infrastructure, offering technologies and tools to best facilitate success.
Henry Weller, product manager, Atlas Vector Search at MongoDB, began by positioning MongoDB’s approach to data management for AI as a developer-first process. MongoDB’s developer data platform simplifies working with data by abstracting infrastructure details, enabling developers to focus their time on more pressing tasks.
“We are trying to provide a wide set of primitives on the same underlying…scalable, secure, multi-cloud platform,” said Weller. “We believe that vector search is another one of these [primitives].”
Utilizing MongoDB as a vector database comes with a variety of advantages, from offering vector embeddings in documents alongside metadata and contextual app data, forming a unified vector and operational data layer, to a single query API with an aggregation framework for advanced vector search filtering and querying. Its seamless developer experience, horizontally, vertically, and independently scalable workloads, and its open and tightly integrated nature makes MongoDB an excellent choice for driving AI success, according to Weller.
Sumit Pal, strategic technology director at Ontotext, posed a question: Why do most data lake efforts for most organizations fail?
Pal pointed to the fact that finding the data itself is becoming increasingly challenging. While many organizations have the right technology in place, the lack of semantics and context forces data into obsolescence.
Fortunately, knowledge graphs solve this exact problem, bridging the gap between information and the knowledge layer, Pal noted. This is especially relevant for “wherever you have data that needs to be connected across different data silos [or] heterogeneous data that encompasses both structured as well as unstructured data. [Ultimately, a] knowledge graph is the most appropriate technology.”
Knowledge graphs are a structured representation of knowledge that capture information in a way that is both human- and machine-readable. By representing data consistently, precisely, and unambiguously, knowledge graphs provide a robust layer of information and all its dependencies.
For this reason, knowledge graphs are the perfect technology to power AI, especially for large language models (LLMs), according to Pal. With a knowledge graph, a company’s LLM produces richer, more relevant responses and trustworthy results, puts AI risk under control and improves quality, as well as reduces integration costs and burden.
“Why can’t we use GenAI [generative AI] as-is today?” asked Noam Liran, CEO at Sightfull. The amount of nuance within analyzing business data—which is often complex and multi-dimensional—makes GenAI implementation a more involved process than “set it and forget it,” Liran explained.
To get trustworthy results from LLMs requires guardrails, Liran noted, which manifests as a semantic layer. Semantic layers act as an abstraction for raw data, providing business context, enforcing approved ontology, and ensuring that the right data is used. By pairing an LLM with a semantic layer, complexities are simplified and better structured through consistent interpretation. Additionally, an LLM underpinned by a semantic layer helps to democratize data to less technical users.
For the full, in-depth webinar, featuring insightful examples, a roundtable discussion, and more, you can view an archived version of the webinar here.