Hallucinations have been a long-standing problem for LLMs. But they are inevitable. Researchers delved into the AI black box to search through the reasons and ultimately, find a solution, but no vain. However, when they looked outside, they found the vector databases can help in reducing the risk of hallucination. 

“Pairing vector databases with LLMs allows for the incorporation of proprietary data, effectively reducing the potential range of responses generated by the database,” said Matt Asay, VP, Developer Relations in an exclusive interaction with AIM at their recently concluded Bengaluru chapter of flagship event MongoDB.local. 

Hallucinations can be controlled through this method. By introducing proprietary data, developers can narrow down the pool of possible responses, significantly reducing the likelihood of hallucinations. However, it was noted that this process requires active programming efforts, and platforms like MongoDB simplify the incorporation of proprietary data into vector databases. 

MongoDB’s vector search capabilities are driving generative AI by transforming diverse data types like text, images, videos, and audio files into numerical vectors, simplifying AI processing and enabling efficient relevance-based searches. MongoDB Atlas with vector search ensures precise information retrieval and personalisation, with dedicated resources for enterprise-scale search workloads.

“While generative AI is important, it’s just one aspect of a broader AI ecosystem where vectors are pivotal,” said Asay. He highlighted that Mongodb’s features, such as the need for streaming data, contribute to the development of smarter applications.

Vectors: One solution for many problems

Vectors play a pivotal role in numerically representing objects and features that were traditionally challenging to store in databases. Vectors offer a solution by enabling the numeric representation of various aspects of the world, which is vital for building effective AI applications.

To facilitate the development of AI applications, Matt explains that they are providing developers with tools to work with vectors, search indexes, metadata, operational data, and documents all in one unified platform. This consolidation of resources is presented as a key strategy to deliver a superior AI experience. 

“We aim to make AI more accessible and approachable for developers, bringing together various data types, including streaming data, vectors, analytics, and more, into a unified platform. This unification is intended to lower the barrier to entry and support developers in building innovative AI applications, aligning with the increasing industry focus on AI development,” he added.

MongoDB for Low Resource Languages?

Recently, Indian IT giant Tech Mahindra introduced Project Indus – a new LLM for Indic languages starting with Hindi. While the majority of AI models like GPT primarily operate in English, support for regional languages is quite limited. To truly democratise AI in India, native language accessibility is indispensable. Project Vaani, Project Bhashini, KissanAI, Jugalbandi are all efforts in this direction. However, there remains immense language data challenges, including the scarcity of diverse and comprehensive datasets, especially in spoken language.

Discussing the potential use of MongoDB in tackling language barriers for AI models,  Asay agreed that MongoDB could indeed handle the vectorization and storage of Hindi language data. However, it was also emphasised that the fundamental issue at hand was the scarcity of language-specific data. MongoDB’s role primarily lay in providing efficient querying and storage capabilities but didn’t directly address the challenge of insufficient language-specific data.

Automobiles Find their Moat in MongoDB

MongoDB plays an important role in the automotive industry by facilitating innovation and enhancing competitiveness. Through MongoDB’s, automotive manufacturers like Toyota, Volvo and Bosch Digital leverage IoT technologies, such as condition monitoring sensors, to expedite informed decision-making. This real-time access to data fosters seamless collaboration among teams, expediting the innovation process.

“Vectorisation plays a crucial role in diagnosing vehicle problems based on engine sounds. This task has always been quite challenging for traditional relational databases,” said Asay.

MongoDB’s Atlas Vector Search transforms sound into queryable data, enabling precise diagnostics. Moreover, repair manuals are similarly vectorized to offer solutions for identified problems. This innovative approach extends to startups utilising vectorization to gauge customer sentiment by processing support call transcripts. Overall, MongoDB empowers automotive manufacturers to seize significant opportunities for groundbreaking products and services, enabling them to remain at the forefront of their industry.

India As a Market 

MongoDB has achieved significant popularity in India, with a rapid customer acquisition rate of about two and a half customers per day in India. Approximately 3,60,000 university registrations out of a total of 1.5 million globally are from India.  The community version of MongoDB has been downloaded roughly 260 million times worldwide, with a substantial portion of these downloads originating in India. The company boasts over 2000 customers in India and is experiencing a remarkable 60% year-on-year growth.

“MongoDB’s platform has gained widespread adoption in the Indian startup ecosystem, spanning sectors like fintech, healthcare, edtech, and more. Our focus areas in India, which include startups, ISVs (Independent Software Vendors), and large enterprises, said Asay.  Some prominent Indian customers include Zomato, cure.fit, Vedantu, and myBillBook. Additionally, MongoDB serves traditional and legacy companies, emphasising the advantages of its document data model, which offers flexibility for developers and scalability, setting it apart from conventional SQL databases.

The post Is MongoDB the Panacea for all LLM Problems?  appeared first on Analytics India Magazine.