Be a part of our each day and weekly newsletters for the newest updates and unique content material on industry-leading AI protection. Be taught Extra
To get the absolute best consequence from an AI question, organizations want the absolute best knowledge.
The reply that many organizations have needed to overcome that problem is retrieval-augmented era (RAG). With RAG, outcomes are grounded in knowledge from a database. Because it seems, although, not all RAG is similar, and really optimizing a database for the absolute best outcomes may be difficult.
Database vendor MongoDB isn’t any stranger to the world of AI or RAG. The corporate’s namesake database is already getting used for RAG, and MongoDB has additionally launched AI purposes improvement initiatives. Whereas the corporate and its customers — such a medical large Novo Nordisk — have had success with gen AI, there’s nonetheless extra to be achieved.
Specifically, hallucination and accuracy continues to be a problem holding some organizations again from getting gen AI into manufacturing. To that finish, MongoDB right now introduced the acquisition of privately-held Voyage AI, which develops superior embedding and retrieval fashions. Voyage raised $20 million in funding in Oct. 2024 in a spherical supported by cloud knowledge large Snowflake. The acquisition will convey Voyage AI’s experience in embedding era and reranking — important parts for AI-powered search and retrieval — straight into MongoDB’s database platform.
“Over the last year, and especially as organizations have tried to think about how they could build AI powered applications, it became increasingly clear that the quality and trust of the applications they build, or the lack thereof, was becoming one of the barriers for applying AI to mission critical use cases,” MongoDB CPO Sahir Azam advised VentureBeat.
What are the challenges of hallucination? Doesn’t RAG clear up them?
The fundamental concept behind RAG is that, as a substitute of merely counting on a data base from educated knowledge, the gen AI engine can get grounded knowledge from a database.
Creating extremely correct RAG is kind of complicated, and there’s nonetheless a possible danger for hallucinations — a problem confronted by MongoDB and its customers. Whereas Azam declined to supply any particular instance or incident the place gen AI RAG failed a person, he did word that accuracy is all the time a priority.
Enhancing accuracy and lowering hallucination includes a number of steps. The primary is to enhance the standard of retrieval (the ‘R’ in RAG).
“In many cases, the retrieval quality is not good enough,” Tengyu Ma, founder and CEO of Voyage AI, advised VentureBeat. “In the retrieval step, if they are not retrieving relevant information, then the retrieval is not very useful, and the large language model (LLM) hallucinates because it has to guess some context.”
The Voyage AI fashions now a part of MongoDB assist enhance RAG in a couple of key methods:
- Area-specific fashions and re-rankers: These are educated on massive quantities of unstructured knowledge from particular verticals, permitting them to higher perceive the terminology and semantics of these domains.
- Customization and fine-tuning: Customers can tremendous tune the retrieval mechanism for distinctive datasets and use instances.
MongoDB’s competitors
MongoDB isn’t the primary or solely vendor to acknowledge the necessity for and worth of getting extremely optimized embedding and re-ranker expertise. In spite of everything, that’s one of many causes Snowflake invested in Voyage AI and is utilizing the corporate’s fashions.
It’s vital to notice that, even after being acquired by MongoDB, Voyage AI’s fashions will nonetheless be accessible to Snowflake and to Voyage AI’s different customers. The large distinction is that Voyage AI will now be more and more built-in into MongoDB’s database platforms.
Immediately integrating superior embedding fashions in a database is an strategy taken by different rival database distributors, as nicely. Again in June 2024, DataStax introduced its personal RAGStack expertise that mixes superior embedding and retrieval fashions.
Azam argued that MongoDB is a bit completely different, although. For one, it’s an operational database, versus an analytical database. Additionally, versus simply offering insights and evaluation, MongoDB helps energy transactions and real-world operations. MongoDB can also be what is called a “document model database,” which has a distinct construction than a standard relational database. That construction doesn’t depend on columns and tables, which aren’t significantly good at representing details about unstructured knowledge (a important aspect for AI purposes).
“We’re the only database technology that combines the management of metadata about a customer’s information, the operations and transactions, which is the heartbeat of what’s happening in the business, as well as the foundation for retrieval — all with a single system,” stated Azam.
Why Voyage AI issues for agentic AI workflows
The necessity for extremely correct embedding and retrieval fashions is being additional accelerated by agentic AI.
“Agentic AI still needs retrieval methods, because an agent cannot make decisions out of context,” stated Ma. “Sometimes, actually multiple retrieval components are used in even one decision.”
Ma famous that Voyage AI is presently engaged on particular fashions which can be extremely personalized for agentic AI use instances. He defined that agentic AI can use various kinds of queries that may nonetheless profit from extra optimization.
As gen AI more and more strikes into operational use instances, the necessity to take away the danger of hallucinations is clearly paramount. Whereas MongoDB has had success with gen AI, Azam expects the mixing of Voyage AI to open new mission important use instances.
“If we can now say, ‘Hey, we can give you well north of 90% accuracy for your applications that today may only, in some cases, get to 30 or 60% accuracy for the results,’ the aperture widens in terms of the types of opportunities people can apply AI to in their software applications,” stated Azam.
