Artificial Intelligence
Building AI-powered Apps with MongoDB
Multi-Agentic Systems in Industry with XMPro & MongoDB Atlas
In 2025, agentic AI applications are no longer pet projects—companies around the world are investing in software to incorporate AI agents into their business workflows. The most common use of an AI agent is to assist with research analysis or writing code. LangChain’s recent survey of over 1000 professionals across multiple industries showed that over 51% have already deployed agents in production, with 60% using the agents for research and summarization tasks. However, leveraging an AI agent for more complex tasks than research and summarization—and implementing them in industrial environments like manufacturing—presents certain challenges. For example, as new technology is introduced into already established companies, the visibility of brownfield deployments increases. This installation and configuration of new hardware or software must coexist with legacy IT systems. And, while it is easy to run an AI agent in a sandboxed environment, it is harder to integrate agents with machines and Operational Technology (OT) systems speaking industrial protocols like Modbus, PROFINET, and BACnet due to existing legacy infrastructure and an accumulation of tech debt. To ensure governance and security in industrial environments, data security policies, regulatory compliance, and governance models are essential. Agent profiles with defined goals, rules, responsibilities, and constraints must be established before agents are deployed. Additionally, addressing real-world constraints—like LLM latency—and strategically selecting use cases and database providers can enhance Al agent effectiveness and optimize response times. What’s more, the successful implementation of AI agents in industrial environments requires a number of foundational elements, including: Flexible data storage and scalability: An agent requires different types of data to function, such as agent profile, short-term memory, and long-term memory. Industrial AI agents require even more types of data, such as time series data from sensors and PLCs. They require efficient and scalable data storage that adapts to the dynamic needs of the environment. Continuous monitoring and analysis: An agent deployed in a manufacturing environment requires real-time observability of ever-changing data generated by the factory. It also needs to keep humans in the loop for any critical decisions that might affect production. High availability: Industrial environments demand near-zero downtime, making system resilience and failover capabilities essential. XMPro joins forces with MongoDB To address these challenges, we are pleased to announce XMPro’s partnership with MongoDB. XMPro offers APEX AI , a low-code control room for creating and managing advanced AI agents for industrial applications. To ensure seamless control over these autonomous agents, XMPro APEX serves as the command center for configuring, monitoring, and orchestrating agent activities, empowering operators to remain in control. Figure 1. XMPro APEX AI platform working with MongoDB Atlas. APEX AI, combined with MongoDB Atlas and MongoDB Atlas Vector Search , addresses a variety of challenges faced by developers when building AI agents for industrial environments. XMPro complements this by seamlessly integrating with industrial equipment such as SCADA systems, PLCs, IoT sensors, and ERPs, enabling continuous monitoring of operations. This integration ensures real-time data acquisition, contextualization, and advanced analytics, transforming raw data into actionable insights. XMPro’s capabilities include condition monitoring, predictive maintenance, anomaly detection, and process optimization, which help reduce downtime and improve operational efficiency while maintaining compliance and safety standards. XMPro’s industrial AI agents rely on memory persistence for contextual decision-making. MongoDB Atlas acts as the database for storing and retrieving agent memories. Using a flexible document database for storing agentic memories enables agents to store different types of data, such as conversational logs, state transitions, and telemetry data, without requiring schema re-design. The capabilities of MongoDB Atlas Vector Search empower APEX AI agents with a retrieval-augmented generation (RAG) tool, which helps to reduce LLM hallucinations. This integration allows agents to access and retrieve verified data, grounding their responses. Having database and vector search tools together in MongoDB Atlas also helps reduce agent latency and speeds up development. APEX AI-enabled multi-agent systems working together in an industrial setting. These context-aware agents can work in tandem, retrieving relevant knowledge stored in MongoDB Atlas to enable meaningful collaboration and better decision-making. XMPro APEX AI also leverages MongoDB Atlas’s robust security and high availability to ensure that agents can securely access and leverage data in real time; features such as role-based access controls, network isolation, encryption in transit and at-rest are key to why this agent-based AI solution is ideal for securing industrial production environments. MongoDB’s highly available and horizontal scalability ensures seamless data access at scale as organizations scale up their APEX AI deployments. Unlocking the future of AI in industrial automation XMPro APEX AI and MongoDB Atlas are a winning combination that paves the way for a new era of industrial automation. By tackling the core challenges of AI agents' deployment in industrial environments, we’re enabling organizations to deploy robust, intelligent, and autonomous industrial AI agents at scale. To learn more about MongoDB’s role in the manufacturing industry, please visit our manufacturing and automotive webpage . Ready to boost your MongoDB skills? Head over to our MongoDB Atlas Learning Hub to start learning today.
VPBank Builds OpenAPI Platform With MongoDB
Open banking is the practice of banks sharing some of their financial data and services to developers for third-party financial service providers through an API. Open banking has accelerated the digitization of the financial services and banking industries. It also helps foster innovation and enhance customer experience by helping create customer-centric, personalized services and experiences. MongoDB has been at the forefront of this revolution. Specifically, MongoDB helps financial institutions worldwide take advantage of OpenAPI . This open-source technology enables an organization’s applications, software, and digital platforms to connect and exchange data with third-party services efficiently and securely. An example is VPBank . One of Vietnam’s largest private banks, it serves over 30 million customers. In 2020, VPBank was the first Vietnamese bank to adopt MongoDB Atlas for OpenAPI. Working with MongoDB, VPBank moved to a microservices architecture , which supported the creation of its own OpenAPI platform and set a new standard for digital banking in Vietnam. Speaking at MongoDB Day in Vietnam in November 2024 , Anh K. Pham, Head of Database Services and Operations for VPBank, shared how MongoDB set up the bank for success with open banking. Migrating from the relational model to the document model Before working with MongoDB, VPBank operated in SQL . The COVID pandemic and the rise of models such as open banking in the early 2020s mandated rapid digitization of banking operations and services. VPBank realized it needed to build the next generation of intelligent banking services to remain competitive. This was not feasible with traditional relational database management systems and the SQL model. VPBank’s primary goal was to harness the power of data and to more efficiently manage unstructured data . This meant switching to an agile architecture based on microservices. “When I was introduced to NoSQL, it made sense,” said Pham. “Data is not always structured. There’s a bunch of different data points here and there, and you can’t make anything of it. But it has to be stored somewhere, it has to be read, and it has to be fed into your applications.” MongoDB Atlas was hosted on Amazon Web Services (AWS) as part of VPBank’s cloud transformation journey. The bank chose MongoDB Atlas for its ability to handle multiple workload types, which had been inadequately supported by its relational databases. These workloads include time series data , event data, real-time analytics, notifications, and big data (like transaction histories, catalog data, and JSON data). Powering 220 microservices with flexibility, scalability, and performance VPBank’s OpenAPI platform consists of over 220 microservices, and it processes more than 100 million transactions per month. By supporting these transactions, MongoDB is ultimately helping VPBank enhance customer experiences and streamline operations. By using MongoDB Atlas, VPBank can better unlock the power of its data to quickly build data-driven applications and services on its microservices architecture. It experienced three substantial benefits by using MongoDB: Flexibility: MongoDB Atlas empowers VPBank to handle complex data, conduct rapid development and iterations, and facilitate efficient API development with BSON. Scalability: MongoDB enables dynamic scaling to handle increasing workloads. Additionally, horizontal scaling distributes data across multiple servers to handle high volumes, spikes in transactions, and API requests. Performance: MongoDB Atlas’s performance capabilities enable VPBank to manage large volumes of data in real time, regardless of acute throughput and latency demands. We have flexibility; we have scalability; we have performance. Those are the main things we want to look at when we’re talking about banking. I need to be flexible. I need to be scalable. I need my performance to be high, because I want my customers to not wait and see if their money is going to go through or not, Ahn K. Pham, Head of Database Services and Operations, VPBank Using OpenShift Container Platform (OCP), VPBank deployed a microservices architecture to run its Open Banking services. “Choosing MongoDB as the modern database was the best choice since it can handle multiple types of data workloads with the performance we needed,” said Pham. Looking to the future VPBank plans to continue its cloud transformation journey. “We’re continuing to migrate our applications from on-premises into the cloud, and we’re continuing to modernize our applications as well,” said Pham. “That means that maybe those other databases that we used to have might be turning into MongoDB databases.” VPBank is also looking at MongoDB to support its AI-driven future: “We really want to focus on AI and data analytics, pulling information from all our customers’ transactions,” explained Pham. “We want to ensure that what we build caters to our 30-plus million customers.” Visit our MongoDB Atlas Learning Hub to boost your MongoDB skills. To learn more about MongoDB for financial services, visit our solutions page .
Transforming News Into Audio Experiences with MongoDB and AI
You wake up, brew your coffee, and start your day with a perfectly tailored podcast summarizing the latest news—delivered in a natural, engaging voice. No manual curation, no human narration, just seamless AI magic. Sounds like the future? It's happening now, powered by MongoDB and generative AI. In 2025, the demand for audio content—particularly podcasts—surged, with 9 million new active listeners in the United States alone, prompting news organizations to seek efficient ways to deliver daily summaries to their audiences. However, automating news delivery has proven to be a challenging task, as media outlets must manage dynamic article data and convert this information into high-quality audio formats at scale. To overcome these hurdles, media organizations can use MongoDB for data storage alongside generative AI for podcast creation, developing a scalable solution for automated news broadcasting. This approach unlocks new AI-driven business opportunities and can attract new customers while strengthening the loyalty of existing ones, contributing to increased revenue streams for media outlets. Check out our AI Learning Hub to learn more about building AI-powered apps with MongoDB. The secret sauce: MongoDB + AI In a news automation solution, MongoDB acts as the system’s backbone, storing news article information as flexible documents with fields like title, content, and publication date—all within a single collection. Alongside this, dynamic elements (such as the number of qualified reads) can be seamlessly integrated into the same document to track content popularity. Moreover, derived insights—e.g., sentiment analysis and key entities—can be generated and enriched through a gen AI pipeline directly within the existing collection. Figure 1. MongoDB data storage for media. This adaptable data structure ensures that the system remains both efficient and scalable, regardless of content diversity or evolving features. As a result, media outlets have created a robust framework to query and extract the latest news and metadata from MongoDB. They can now integrate AI with advanced language models to transform this information into an audio podcast. With this foundation in place, let's examine why MongoDB is well-suited for implementing AI-driven applications. Why MongoDB is the perfect fit News data is inherently diverse, with each article containing a unique mix of attributes, including main content fields (e.g. id, title, body, date, imageURL), calculated meta data (e.g. read count), generated fields with the help of GenAI (e.g. keywords, sentiment) and embeddings for semantic/vector search. Some of these elements originate from publishers, while others emerge from user interactions or AI-driven analysis. MongoDB’s flexible document model accommodates all these attributes—whether predefined or dynamically generated, within a single, adaptable structure. This eliminates the rigidity of traditional databases and ensures that the system evolves seamlessly alongside the data it manages. What’s more, speed is critical in news automation. By storing complete, self-contained documents, MongoDB enables rapid retrieval and processing without the need for complex joins. This efficiency allows articles to be enriched, analyzed, and transformed into audio content in near real-time. And scalability is built in. Whether handling a small stream of updates or processing vast amounts of constantly changing data, MongoDB’s distributed architecture ensures high availability and seamless growth, making it ideal for large-scale media applications. Last but hardly least, developers benefit from MongoDB’s agility. Without the constraints of fixed schemas, new data points—whether from evolving AI models, audience engagement metrics, or editorial enhancements—can be integrated effortlessly. This flexibility allows teams to experiment, iterate, and scale without friction, ensuring that the system remains future-proof as news consumption evolves. Figure 2. MongoDB benefits for AI-driven applications. Bringing news to life with generative AI Selecting MongoDB for database storage is just the beginning; the real magic unfolds when text meets AI-powered speech synthesis. In our labs, we have experimented with Google’s NotebookLM model to refine news text, ensuring smooth narration with accurate intonation and pacing. Putting all these pieces together, the diagram below illustrates the workflow for automating AI-based news summaries into audio conversions. Figure 3. AI-based text-to-audio conversion architecture. The process begins with a script that retrieves relevant news articles from MongoDB, using the Aggregation Framework and Vector Search to ensure semantic relevance. These selected articles are then passed through an AI-powered pipeline, where they are condensed into a structured podcast script featuring multiple voices. Once the script is refined, advanced text-to-speech models transform it into high-quality audio, which is stored as a .wav file. To optimize delivery, the generated podcast is cached, ensuring seamless playback for users on demand. The result? A polished, human-like narration, ready for listeners in MP3 format. Thanks to this implementation, media outlets can finally let go of the robotic voices of past automations. Instead, they can now deliver a listening experience to their customers that's human, engaging, and professional. The future of AI-powered news consumption This system isn’t just a technological innovation; it’s a revolution in how we consume news. By combining MongoDB’s efficiency with AI’s creative capabilities, media organizations can deliver personalized, real-time news summaries without human intervention. It’s faster, smarter, and scalable—ushering in a new era of automated audio content. Want to build the next-gen AI-powered media platform? Start with MongoDB and let your content speak for itself! To learn more about integrating AI into media systems using MongoDB, check out the following resources to guide your next steps: The MongoDB Solutions Library: Gen AI-powered video summarization The MongoDB Blog: AI-Powered Media Personalization: MongoDB and Vector Search
GraphRAG with MongoDB Atlas: Integrating Knowledge Graphs with LLMs
A key challenge AI developers face is providing context to large language models (LLMs) to build reliable AI-enhanced applications; retrieval-augmented generation (RAG) is widely used to tackle this challenge. While vector-based RAG, the standard (or baseline) implementation of retrieval-augmented generation, is useful for many use cases, it is limited in providing LLMs with reasoning capabilities that can understand relationships between diverse concepts scattered throughout large knowledge bases. As a result, the accuracy of vector RAG-enhanced LLM outputs in applications can disappoint—and even mislead—end users. Now generally available, MongoDB Atlas ’ new LangChain integration for GraphRAG—a variation of RAG architecture that integrates a knowledge graph with LLMs—can help address these limitations. GraphRAG: Connecting the dots First, a short explanation of knowledge graphs: a knowledge graph is a structured representation of information in which entities (such as people, organizations, or concepts) are connected by relationships. Knowledge graphs work like maps, and show how different pieces of information relate to each other. This structure helps computers understand connections between facts, answer complex questions, and find relevant information more easily. Traditional RAG applications split knowledge data into chunks, vectorize them into embeddings, and then retrieve chunks of data through semantic similarity search; GraphRAG builds on this approach. But instead of treating each document or chunk as an isolated piece of information, GraphRAG considers how different pieces of knowledge are connected and relate to each other through a knowledge graph. Figure 1. Embedding-based vector search vs. entity-based graph search. GraphRAG improves RAG architectures in three ways: First, GraphRAG can improve response accuracy . Integrating knowledge graphs into the retrieval component of RAG has shown significant improvements in multiple publications. For example, benchmarks in the AWS investigation, “ Improving Retrieval Augmented Generation Accuracy with GraphRAG ” demonstrated nearly double the correct answers compared to traditional embedding-based RAG. Also, embedding-based methods rely on numerical vectors and can make it difficult to interpret why certain chunks are related. Conversely, a graph-based approach provides a visual and auditable representation of document relationships. Consequently, GraphRAG offers more explainability and transparency into retrieved information for improved insight into why certain data is being retrieved. These insights can help optimize data retrieval patterns to improve accuracy. Finally, GraphRAG can help answer questions that RAG is not well-suited for—particularly when understanding a knowledge base's structure, hierarchy, and links is essential . Vector-based RAG struggles in these cases because breaking documents into chunks loses the big picture. For example, prompts like “What are the themes covered in the 2025 strategic plan?” are not well handled. This is because the semantic similarity between the prompt, with keywords like “themes,” and the actual themes in the document may be weak, especially if they are scattered across different sections. Another example prompt like, “What is John Doe’s role in ACME’s renewable energy projects?” presents challenges because if the relationships between the person, the company, and the related projects are mentioned in different places, it becomes difficult to provide accurate responses with vector-based RAG. Traditional vector-based RAG can struggle in cases like these because it relies solely on semantic similarity search. The logical connections between different entities—such as contract clauses, legal precedents, financial indicators, and market conditions—are often complex and lack semantic keyword overlap. Making logical connections across entities is often referred to as multi-hop retrieval or reasoning in GraphRAG. However, GraphRAG has its own limitations, and is use-case dependent to achieve better accuracy than vector-based RAG: It introduces an extra step: creating the knowledge graph using LLMs to extract entities and relationships. Maintaining and updating the graph as new data arrives becomes an ongoing operational burden. Unlike vector-based RAG, which requires embedding and indexing—a relatively lightweight and fast process—GraphRAG depends on a large LLM to accurately understand, map complex relationships, and integrate them into the existing graph. The added complexity of graph traversal can lead to response latency and scalability challenges as the knowledge base grows. Latency is closely tied to the depth of traversal and the chosen retrieval strategy, both of which must align with the specific requirements of the application. GraphRAG introduces additional retrieval options . While this allows developers more flexibility in the implementation, it also adds complexity. The additional retrieval options include keyword and entity-based retrieval, semantic similarity on the first node, and more. MongoDB Atlas: A unified database for operational data, vectors, and graphs MongoDB Atlas is perfectly suited as a unified database for documents, vectors, and graphs. As a unified platform, it’s ideal for powering LLM-based applications with vector-based or graph-based RAG. Indeed, adopting MongoDB Atlas eliminates the need for point or bolt-on solutions for vector or graph functionality, which often introduce unnecessary complexity, such as data synchronization challenges that can lead to increased latency and potential errors. The unified approach offered by MongoDB Atlas simplifies the architecture and reduces operational overhead, but most importantly, it greatly simplifies the development experience. In practice, this means you can leverage MongoDB Atlas' document model to store rich application data, use vector indexes for similarity search, and model relationships using document references for graph-like structures. Implementing GraphRAG with MongoDB Atlas and LangChain Starting from version 0.5.0, the langchain-mongodb package introduces a new class to simplify the implementation of a GraphRAG architecture. Figure 2. GraphRAG architecture with MongoDB Atlas and LangChain First, it enables the automatic creation of a knowledge graph. Under the hood, it uses a specific prompt sent to an LLM of your choice to extract entities and relationships, structuring the data to be stored as a graph in MongoDB Atlas. Then, it sends a query to the LLM to extract entities and then searches within the graph to find connected entities, their relationships, and associated data. This information, along with the original query, then goes back to the LLM to generate an accurate final response. MongoDB Atlas’ integration in LangChain for GraphRAG follows an entity-based graph approach. However, you can also develop and implement your own GraphRAG with a hybrid approach using MongoDB drivers and MongoDB Atlas’ rich search and aggregation capabilities. Enhancing knowledge retrieval with GraphRAG GraphRAG complements traditional RAG methods by enabling deeper understanding of complex, hierarchical relationships, supporting effective information aggregation and multi-hop reasoning. Hybrid approaches that combine GraphRAG with embedding-based vector search further enhance knowledge retrieval, making them especially effective for advanced RAG and agentic systems. MongoDB Atlas’ unified database simplifies RAG implementation and its variants, including GraphRAG and other hybrid approaches, by supporting documents, vectors, and graph representations in a unified data model that can seamlessly scale from prototype to production. With robust retrieval capabilities ranging from full-text and semantic search to graph search, MongoDB Atlas provides a comprehensive solution for building AI applications. And its integration with proven developer frameworks like LangChain accelerates the development experience—enabling AI developers to build more advanced and efficient retrieval-augmented generation systems that underpin AI applications. Ready to dive into GraphRAG? Learn how to implement it with MongoDB Atlas and LangChain. Head over to the Atlas Learning Hub to boost your MongoDB skills and knowledge.
Next-Generation Mobility Solutions with Agentic AI and MongoDB Atlas
Driven by advancements in vehicle connectivity, autonomous systems, and electrification, the automotive and mobility industry is currently undergoing a significant transformation. Vehicles today are sophisticated machines, computers on wheels, that generate massive amounts of data, driving demand for connected and electric vehicles. Automotive players are embracing artificial intelligence (AI), battery electrical vehicles (BEVs), and software-defined vehicles (SDVs) to maintain their competitive advantage. However, managing fleets of connected vehicles can be a challenge. As cars get more sophisticated and are increasingly integrated with internal and external systems, the volume of data they produce and receive greatly increases. This data needs to be stored, transferred, and consumed by various downstream applications to unlock new business opportunities. This will only grow: the global fleet management market is projected to reach $65.7 billion by 2030, growing at a rate of almost 10.8% annually. A 2024 study conducted by Webfleet showed that 32% of fleet managers believe AI and machine learning will significantly impact fleet operations in the coming years; optimizing route planning and improving driver safety are the two most commonly cited use cases. As fleet management software providers continue to invest in AI, the integration of agentic AI can significantly help with things like route optimization and driver safety enhancement. For example, AI agents can process real-time traffic updates and weather conditions to dynamically adjust routes, ensuring timely deliveries while advising drivers on their car condition. This proactive approach contrasts with traditional reactive methods, improving vehicle utilization and reducing operational and maintenance costs. But what are agents? In short, they are operational applications that attempt to achieve goals by observing the world and acting upon it using the data and tools the application has at its disposal. The term "agentic" denotes having agency, as AI agents can proactively take steps to achieve objectives without constant human oversight. For example, rather than just reporting an anomaly based on telemetry data analysis, an agent for a connected fleet could autonomously cross-check that anomaly against known issues, decide whether it's critical or not, and schedule a maintenance appointment all on its own. Why MongoDB for agentic AI Agentic AI applications are dynamic by nature as they require the ability to create a chain of thought, use external tools, and maintain context across their entire workflow. These applications generate and consume diverse data types, including structured and unstructured data. MongoDB’s flexible document model is uniquely suited to handle both structured and unstructured data as vectors. It allows all of an agent’s context, chain-of-thought, tools metadata, and short-term and long-term memory to be stored in a single database. This means that developers can spend more time on innovation and rapidly iterate on agent designs without being constrained by rigid schemas of a legacy relational database. Figure 1. Major components of an AI agent. Figure 1 shows the major components of an AI agent. The agent will first receive a task from a human or via an automated trigger, and will then use a large language model (LLM) to generate a chain of thought or follow a predetermined workflow. The agent will use various tools and models during its run and store/retrieve data from a memory provider like MongoDB Atlas . Tools: The agent utilizes tools to interact with the environment. This can contain API methods, database queries, vector search, RAG application, anything to support the model Models: can be a large language model (LLM), vision language model (VLM), or a simple supervised machine learning model. Models can be general purpose or specialized, and agents may use more than one. Data: An agent requires different types of data to function. MongoDB’s document model allows you to easily model all of this data in one single database. An agentic AI spans a wide range of functional tools and context. The underlying data structures evolve throughout the agentic workflow and as an agent uses different tools to complete a task. It also builds up memory over time. Let us list down the typical data types you will find in an agentic AI application. Data types: Agent profile: This contains the identity of the agent. It includes instructions, goals and constraints. Short-term memory: This holds temporary, contextual information—recent data inputs or ongoing interactions—that the agent uses in real-time. For example, short-term memory could store sensor data from the last few hours of vehicle activity. In certain agentic AI frameworks like Langgraph, short term memory is implemented through a checkpointer. The checkpointer stores intermediate states of the agent’s actions and/or reasoning. This memory allows the agent to seamlessly pause and resume operations. Long-term memory: This is where the agent stores accumulated knowledge over time. This may include patterns, trends, logs and historical recommendations and decisions. By storing each of these data types into rich, nested documents in MongoDB, AI developers can create a single-view representation of an agent’s state and behavior. This enables fast retrieval and simplifies development. In addition to the document model advantage, building agentic AI solutions for mobility requires a robust data infrastructure. MongoDB Atlas offers several key advantages that make it an ideal foundation for these AI-driven architectures. These include: Scalability and flexibility: Connected Car platforms like fleet management systems need to handle extreme data volumes and variety. MongoDB Atlas is proven to scale horizontally across cloud clusters, letting you ingest millions of telemetry events per minute and store terabytes of telemetry data with ease. For example, the German company ZF uses MongoDB to process 90,000 vehicle messages per minute (over 50 GB of data per day) from hundreds of thousands of connected cars. The flexibility of the document model accelerates development and ensures your data model stays aligned with the real-world entities it represents. Built-in vector search: AI agents require a robust set of tools to work with. One of the most widely used tools is vector search, which allows agents to perform semantic searches on unstructured data like driver logs, error codes descriptions, and repair manuals. MongoDB Atlas Vector Search allows you to store and index high-dimensional vectors alongside your documents and to perform semantic search over unstructured data. In practice, this means your AI embeddings live right next to the relevant vehicle telemetry and operational data in the database, simplifying architectures for use cases like the connected car incident advisor, in which a new issue can be matched against past issues before passing contextual information to the LLM. For more, check out this example of how an automotive OEM leverages vector search for audio based diagnostics with MongoDB Atlas Vector Search. Time series collections and real-time data processing: MongoDB Atlas is designed for real-time applications. It provides time series collections for connected car telemetry data storage, change streams, and triggers that can react to new data instantly. This is crucial for agentic AI feedback loops, where ongoing data ingestion and learning are happening continuously. Best-in-class embedding models with Voyage AI: In early 2025, MongoDB acquired Voyage AI , a leader in embedding and reranking models. Voyage AI embedding models are currently being integrated into MongoDB Atlas, which means developers will no longer need to manage external embedding APIs, standalone vector stores, or complex search pipelines. AI retrieval will be built into the database itself, making semantic search, vector retrieval, and ranking as seamless as traditional queries. This will reduce the time required for developing agentic AI applications. Agentic AI in action: Connected fleet incident advisor Figure 2 shows a list of use cases in the Mobility sector, sorted by various capabilities that an agent might demonstrate. AI agents excel at managing multi-step tasks via context management across tasks, they automate repetitive tasks better than Robotic process automation (RPA), and they demonstrate human-like reasoning by revisiting and revising past decisions. These capabilities enable a wide range of applications both during the manufacturing of a vehicle and while it's on the road, connected and sending telemetry. We will review a use case in detail below, and will see how it can be implemented using MongoDB Atlas, LangGraph, Open AI, and Voyage AI. Figure 2. Major use cases of agentic AI in the mobility and manufacturing sectors. First, the AI agent connects to traditional fleet management software and supports the fleet manager in diagnosing and advising the drivers. This is an example of a multi-step diagnostic workflow that gets triggered when a driver submits a complaint about the vehicle's performance (for example, increased fuel consumption). Figure 3 shows the sequence diagram of the agent. Upon receiving the driver complaint, it creates a chain of thought that follows a multi-step diagnostic workflow where the system ingests vehicle data such as engine codes and sensor readings, generates embeddings using the Voyage AI voyage-3-large embedding model, and performs a vector search using MongoDB Atlas to find similar past incidents. Once relevant cases are identified, those–along with selected telemetry data–are passed to OpenAI gpt-4o LLM to generate a final recommendation for the driver (for example, to pull off immediately or to keep driving and schedule regular maintenance). All data, including telemetry, past issues, session logs, agent profiles, and recommendations are stored in MongoDB Atlas, ensuring traceability and the ability to refine diagnostics over time. Additionally, MongoDB Atlas is used as a checkpointer by LangGraph, which defines the agent's workflow. Figure 3. Sequence diagram for a connected fleet advisor agentic workflow. Figure 4 shows the agent in action, from receiving an issue to generating a recommendation. So by leveraging MongoDB’s flexible data model and powerful Vector Search capabilities, we can agentic AI can transform fleet management through predictive maintenance and proactive decision-making. Figure 4. The connected fleet advisor AI agent in action. To set up the use case shown in this article, please visit our GitHub repository . And to learn more about MongoDB’s role in the automotive industry, please visit our manufacturing and automotive webpage . Want to learn more about why MongoDB is the best choice for supporting modern AI applications? Check out our on-demand webinar, “ Comparing PostgreSQL vs. MongoDB: Which is Better for AI Workloads? ” presented by MongoDB Field CTO, Rick Houlihan.
MongoDB Powers M-DAQ’s Anti-Money Laundering Compliance Platform
Founded and headquartered in Singapore, M-DAQ Global is a fintech powerhouse providing seamless cross-border transactions for businesses worldwide. M-DAQ’s comprehensive suite of foreign exchange, collections, and payments solutions help organizations of all sizes navigate the complexities of global trade, offering FX clarity, certainty, and payment mobility. M-DAQ also offers AI-powered services like Know Your Business (KYB), onboarding, and advanced risk management tools. Amidst ever-evolving requirements, these enable business transactions across borders with ease, while staying compliant. One of M-DAQ's most innovative solutions, CheckGPT , is an AI-powered platform designed to streamline Anti-Money Laundering (AML) compliance. It was built on MongoDB Atlas , providing a strong foundation for designing multitenant data storage. This approach ensures that each client has a dedicated database, effectively preventing any data co-mingling. Traditional AML processes often involve tedious, time-consuming tasks, from document review, to background checks, to customer onboarding. By building CheckGPT, M-DAQ’s aim was to change this paradigm, and to leverage AI to automate (and speed) these manual processes. Today, CheckGPT allows businesses to process onboarding 30 times faster than traditional human processing. The platform also leverages MongoDB Atlas’s native Vector Search capabilities to power intelligent semantic searches across unstructured data. The challenge: Managing unstructured, sensitive data, and performing complex searches One of CheckGPT’s priorities was to improve processes around collecting, summarizing, and analyzing data, while flagging potential risks to customers quickly and accurately. Considering the vast number and complexity of data sets its AI platform had to handle, and the strict regulatory landscape the company operates in, it was crucial that M-DAQ chose a robust database. CheckGPT needed a database that could efficiently and accurately handle unstructured data, and adapt rapidly as the data evolved. The database also had to be highly secure; to function, the AI tool would have to handle highly sensitive data, and would need to be used by companies operating in highly regulated industries. Finally, CheckGPT was looking for the ability to perform complex, high-dimensional searches to power a wide range of complex searches and real-time information analysis. MongoDB Atlas: A complete platform with unique features According to M-DAQ, there are many benefits of using MongoDB Atlas’ document model: Flexibility: MongoDB Atlas’s document model accommodates the evolving nature of compliance data, providing the flexibility needed to manage CheckGPT's dynamic data structures, such as onboarding documents and compliance workflows. Security and performance: The MongoDB Atlas platform also ensures that data remains secure throughout its lifecycle. M-DAQ was able to implement a multi-tenancy architecture that securely isolates data across its diverse client base. This ensures that the platform can handle varying compliance demands while maintaining exceptional performance, giving M-DAQ’s customers the confidence that the AML processes handled by CheckGPT are compliant with stringent regulatory standards. Vector search capabilities: MongoDB Atlas provides a unified development experience. Particularly, MongoDB Atlas Vector Search enables real-time searches across a vast amount of high-dimensional datasets. This makes it easier to verify documents, conduct background checks, and continuously monitor customer activity, ensuring fast and accurate results during AML processes. “AI, together with the flexibility of MongoDB, has greatly impacted CheckGPT, enabling us to scale operations and automate complex AML compliance processes,” said Andrew Marchen, General Manager, Payments and Co-founder, Wallex at M-DAQ Global. “This integration significantly reduces onboarding time, which typically took between 4-8 hours to three days depending on the document’s complexity, to less than 10 minutes. With MongoDB, M-DAQ is able to deliver faster and more accurate results while meeting customer needs in a secure and adaptable environment." The future of CheckGPT, powered by MongoDB M-DAQ believes that AI and data-driven technologies and tools will continue to play a central role in automating complex processes. By employing AI, M-DAQ aims to improve operational efficiency, enhance customer experiences, and scale rapidly—while maintaining high service standards. MongoDB’s flexibility and multi-cloud support will be key as M-DAQ plans to use single/multi-cluster and multi-region capabilities in the future. M-DAQ aims to explore additional features that could enhance CheckGPT's scalability and performance. The company, for example, plans to expand its use of MongoDB for future projects involving automating complex processes like compliance, onboarding, and risk management in 2025. Learn more about CheckGPT on their site . Visit our product page to learn more about MongoDB Atlas. Get started with MongoDB Atlas Vector Search today with our Atlas Vector Search Quick Start guide .
Building Gen AI with MongoDB & AI Partners | February 2025
February was big for MongoDB—and, more importantly, for anyone looking to build AI applications that deliver highly accurate, relevant information (in other words, for everyone building AI apps). MongoDB announced the acquisition of Voyage AI , a pioneer in state-of-the-art embedding and reranking models that power next-generation AI applications. Because generative AI is by nature probabilistic, models can “hallucinate”, and generate false or misleading information. This can lead to serious risks, especially in cases or industries (e.g., financial services) where accurate information is paramount. To address this, organizations building AI apps need high-quality retrieval; they need to trust that the most relevant information is extracted from their data with precision. Voyage AI’s advanced embedding and reranking models enable applications to extract meaning from highly specialized and domain-specific text and unstructured data. With roots at Stanford and MIT, Voyage AI’s world-class team is trusted by AI innovators like Anthropic, LangChain, Harvey, and Replit. Integrating Voyage AI’s technology with MongoDB will enable organizations to easily build trustworthy, AI-powered applications by offering highly accurate and relevant information retrieval deeply integrated with operational data. For more, check out MongoDB CEO Dev Ittycheria’s blog post about Voyage AI , and what this means for developers and businesses (in short, delivering high-quality results at scale). Onward! P.S. If you’re in Vegas for HumanX this week, stop by booth 412 to say hi to MongoDB! Welcoming new AI and tech partners The Voyage AI news was hardly the only exciting development last month. In February 2025, MongoDB welcomed three new AI and tech partners that offer product integrations with MongoDB. Read on to learn more about each great new partner! CopilotKit Seattle-based CopilotKit provides open source infrastructure for in-app AI copilots. CopilotKit helps organizations build production-ready copilots and agents effortlessly. “We’re excited to be partnering with MongoDB to help companies build best-in-class copilots that leverage RAG & take action based on internal data,” said Uli Barkai, Co-Founder and Chief Marketing Officer at CopilotKit. “MongoDB made it dead simple to build a scalable vector database with operational data. This collaboration enables developers to easily ship production-grade RAG applications.” Varonis Varonis is the leader in data security, protecting data wherever it lives—across SaaS, IaaS, and hybrid cloud environments. Varonis’ cloud-native Data Security Platform continuously discovers and classifies critical data, removes exposures, and detects advanced threats with AI-powered automation. “Varonis’s mission is to protect data wherever it lives,” said David Bass, Executive Vice President of Engineering and Chief Technology Officer at Varonis. “We are thrilled to further advance our mission by offering AI-powered data security and compliance for MongoDB, the database of choice for high-performance application and AI development. With this integration, joint customers can automatically discover and classify sensitive data, detect abnormal activities, secure AI data pipelines, and prevent data leaks.” Xlrt Xlrt is an automated insight-generation platform that enables financial institutions to create innovative financial credit products at scale by simplifying the financial spreading process. “We are excited to partner with MongoDB Atlas to transform AI-driven financial workflows,” said Rupesh Chaudhuri, Chief Operating Officer and Co-Founder of Xlrt. “XLRT.ai leverages agentic AI, combining graph-based contextualization, vector search, and LLMs to redefine data-driven decision-making. With MongoDB's robust NoSQL and vector search capabilities, we’re delivering unparalleled efficiency, accuracy, and scalability in automating financial processes.” To learn more about building AI-powered apps with MongoDB, check out our AI Learning Hub and stop by our Partner Ecosystem Catalog to read about our integrations with MongoDB’s ever-evolving AI partner ecosystem. And visit the MongoDB AI Applications Program (MAAP) page to learn how MongoDB and the MAAP ecosystem helps organizations build applications with advanced AI capabilities.
ORiGAMi: A Machine Learning Architecture for the Document Model
The document model has proven to be the optimal paradigm for modern application schemas. At MongoDB, we've long understood that semi-structured data formats like JSON offer superior expressiveness compared to traditional tabular and relational representations. Their flexible schema accommodates dynamic and nested data structures, naturally representing complex relationships between data entities. However, the machine learning (ML) community has faced persistent challenges when working with semi-structured formats. Traditional ML algorithms, as implemented in popular libraries like scikit-learn and pandas , operate on the assumption of fixed-dimensional tabular data consisting of rows and columns. This fundamental mismatch forces data scientists to manually convert JSON documents into tabular form—a time-consuming process that requires significant domain expertise. Recent advances in natural language processing (NLP) demonstrate the power of Transformers in learning from unstructured data but their application to semi-structured data has been under-studied. To bridge this gap, MongoDB's ML research group has developed a novel Transformer-based architecture designed for supervised learning on semi-structured data (e.g., JSON data in a document model database). We call this new architecture ORiGAMi (Object Representation through Generative, Autoregressive Modelling), and we're excited to make it available to the community at github.com/mongodb-labs/origami . It includes components that make training a Transformer model feasible on datasets entailing as few as 200 labeled samples. By combining this data efficiency with the flexibility of Transformers, ORiGAMi enables prediction directly from semi-structured documents, without the cumbersome flattening and manual feature extraction required for tabular data representation. You can read more about our model on arXiv . Technical innovation The key insight behind ORiGAMi lies in its tokenization strategy: documents are transformed into sequences of key-value pairs and special structural tokens that encode nested types like arrays and subdocuments: These token sequences serve as input to the Transformer model trained to predict the next token given a portion of the document, similar to how large language models (LLMs) are trained on text tokens. What’s more, our modifications to the standard Transformer architecture include guardrails to ensure that the model only generates valid, well-formed documents, and a novel position encoding strategy that respects the order invariance of key/value pairs in JSON. These modifications also allow for much smaller models compared to LLMs, which can thus be trained on consumer hardware in minutes to hours depending on dataset size and complexity, versus days to weeks for LLMs. By reformulating classification as a next-token prediction task, ORiGAMi can predict any field within a document, including complex types like arrays and nested subdocuments. This unified approach eliminates the need for separate models or preprocessing pipelines for different prediction tasks. Example use case Our initial focus has been supervised learning: training models from labeled data to make predictions on unseen documents. Let's explore a practical example of user segmentation. Consider a collection where each document represents a user profile, containing both simple fields and complex nested structures: { "_id": "user_7842", "email": "sarah.chen@example.com", "signup_date": "2024-01-15", "device_history": [ { "device": "mobile_ios", "first_seen": "2024-01-15", "last_seen": "2024-02-11" }, { "device": "desktop_chrome", "first_seen": "2024-01-16", "last_seen": "2024-02-10" } ], "subscription": { "plan": "pro", "billing_cycle": "annual", "features_used": ["analytics", "api_access", "team_sharing"], "usage_metrics": { "storage_gb": 45.2, "api_calls_per_day": 1250, "active_projects": 8 } }, "user_segment": "enterprise_power_user" // <-- target field } Suppose you want to automatically classify users into segments like "enterprise_power_user", "smb_growth", or "early_stage_startup" based on their behavior and characteristics. Some documents in your collection already have correct labels, perhaps assigned through manual analysis or customer interviews. Traditional ML approaches would require flattening this rich document structure, leading to very sparse tables and potentially losing important hierarchical relationships. With ORiGAMi, you can: Train directly on the raw documents with existing labels Preserve the full context of nested structures and arrays Make predictions for the "user_segment" field on new users immediately after signup Update predictions as user behavior evolves without rebuilding feature pipelines Getting started with ORiGAMi We're excited to be open-sourcing ORiGAMi ( github.com/mongodb-labs/origami ) and you can read more about our model on arXiv . We've also included a command-line interface that lets users make predictions without writing any code. Training a model is as simple as pointing ORiGAMi to your MongoDB collection: origami train <mongo-uri> -d app -c users Once trained, you can generate predictions and seamlessly integrate them back into your MongoDB workflow. For example, to predict user segments for new signups (from the analytics.signups collection ) and write the resulting predictions back to MongoDB to an analytics.predicted collection: origami predict <mongo-uri> -d analytics -c signups --target user_segment --json | mongoimport -d analytics -c predicted For those looking to dive deeper, we've also included several Jupyter notebooks in the repository that demonstrate advanced features and customization options. Model performance can be improved by adjusting the hyperparameters. We're just scratching the surface of what's possible with document-native machine learning, and have many more use cases in mind. We invite you to explore the repository, contribute to the project, and share how you use ORiGAMi to solve real-world problems. Head over to the ORiGAMi github repo , play around with it, and tell us about new ways of applying it and problems it’s well-suited to solving.
AI-Powered Java Applications With MongoDB and LangChain4j
MongoDB is pleased to introduce its integration with LangChain4j , a popular framework for integrating large language models (LLMs) into Java applications. This collaboration simplifies the integration of MongoDB Atlas Vector Search into Java applications for building AI applications. The advent of generative AI has opened up many new possibilities for developing novel applications. These advancements have led to the development of AI frameworks that simplify the complexities of orchestrating and integrating LLMs and the various components of the AI stack , where MongoDB plays a key role as an operational and vector database. Simplifying AI development for Java The first AI frameworks to emerge were developed for Python and JavaScript, which were favored by early AI developers. However, Java remains widespread in enterprise software. This has led to the development of LangChain4j to address the needs of the Java ecosystem. While largely inspired by LangChain and other popular AI frameworks, LangChain4j is independently developed. As with other LLM frameworks, LangChain4j offers several advantages for developing AI systems and applications by providing: A unified API for integrating LLM providers and vector stores. This enables developers to adopt a modular approach with an interchangeable stack while ensuring a consistent developer experience. Common abstractions for LLM-powered applications, such as prompt templating, chat memory management, and function calling, offering ready-to-use building blocks for common AI applications like retrieval-augmented generation (RAG) and agents. Powering RAG and agentic systems with MongoDB and LangChain4j MongoDB worked with the LangChain4j open-source community to integrate MongoDB Atlas Vector Search into the framework, enabling Java developers to develop AI-powered applications from simple RAG to agentic applications. In practice, this means developers can now use the unified LangChain4j API to store vector embeddings in MongoDB Atlas and use Atlas Vector Search capabilities for retrieving relevant context data. These capabilities are essential for enabling RAG pipelines, where private, often enterprise data is retrieved based on relevancy and combined with the original prompt to get more accurate results in LLM-based applications. LangChain4j supports various levels of RAG, from basic to advanced implementations, making it easy to prototype and experiment before customizing and scaling your solution to your needs. A basic RAG setup with LangChain4j typically involves loading and parsing unstructured data from documents stored locally or on remote services like Amazon S3 or Azure Storage using the Document API. The process then transforms and splits the data, then embeds it to capture the semantic meaning of the content. For more details, check out the documentation on core RAG APIs . However, real-world use cases often demand solutions with advanced RAG and agentic systems. LangChain4j optimizes RAG pipelines with predefined components designed to enhance accuracy, latency, and overall efficiency through techniques like query transformation, routing, content aggregation, and reranking. It also supports AI agent implementation through dedicated APIs, such as AI Services and Tools , with function calling and RAG integration, among others. Learn more about the MongoDB Atlas Vector Search integration in LangChain4j’s documentation . MongoDB’s dedication to providing the best developer experience for building AI applications across different ecosystems remains strong, and this integration reinforces that commitment. We will continue strengthening our integration with LLM frameworks enabling developers to build more-innovative AI applications, agentic systems, and AI agents. Ready to start building AI applications with Java? Learn how to create your first RAG system by visiting our tutorial: How to Make a RAG Application With LangChain4j .
Why Vector Quantization Matters for AI Workloads
Key takeaways As vector embeddings scale into millions, memory usage and query latency surge, leading to inflated costs and poor user experience. By storing embeddings in reduced-precision formats (int8 or binary), you can dramatically cut memory requirements and speed up retrieval. Voyage AI's quantization-aware embedding models are specifically tuned to handle compressed vectors without significant loss of accuracy. MongoDB Atlas streamlines the workflow by handling the creation, storage, and indexing of compressed vectors, enabling easier scaling and management. MongoDB is built for change, allowing users to effortlessly scale AI workloads as resource demands evolve. Organizations are now scaling AI applications from proofs of concept to production systems serving millions of users. This shift creates scalability, latency, and resource challenges for mission-critical applications leveraging recommendation engines, semantic search, and retrieval-augmented generation (RAG) systems. At scale, minor inefficiencies compound and become major bottlenecks, increasing latency, memory usage, and infrastructure costs. This guide explains how vector quantization enables high-performance, cost-effective AI applications at scale. The challenge: Scaling vector search in production Let’s start by considering a modern voice assistance platform that combines semantic search with natural language understanding. During development, the system only needs to process a few hundred queries per day, converting speech to text and matching the resulting embeddings against a modest database of responses. The initial implementation is straightforward: each query generates a 32-bit floating-point embedding vector that's matched against a database of similar vectors using cosine similarity. This approach works smoothly in the prototype phase—response times are quick, memory usage is manageable, and the development team can focus on improving accuracy and adding features. However, as the platform gains traction and scales to processing thousands of queries per second against millions of document embeddings, the simple approach begins to break down. Each incoming query now requires loading massive amounts of high-precision floating-point vectors into memory, computing similarity scores across an exponentially larger dataset, and maintaining increasingly complex vector indexes for efficient retrieval. Without proper optimization, the system struggles as memory usage balloons, query latency increases, and infrastructure costs spiral upward. What started as a responsive, efficient prototype has become a bottleneck production system that struggles to maintain its performance requirements while serving a growing user base. The key challenges are: Loading high-precision 32-bit floating-point vectors into memory Computing similarity scores across massive embedding collections Maintaining large vector indexes for efficient retrieval Which can lead to critical issues like: High memory usage as vector databases struggle to keep float32 embeddings in RAM Increased latency as systems process large volumes of high-precision data Growing infrastructure costs as organizations scale their vector operations Reduced query throughput due to computational overhead AI workloads with tens or hundreds of millions of high-dimensional vectors (e.g., 80M+ documents at 1536 dimensions) face soaring RAM and CPU requirements. Storing float32 embeddings for these workloads can become prohibitively expensive. Vector quantization: A path to efficient scaling The obvious question is: How can you maintain the accuracy of your recommendations, semantic matches, and search queries, while drastically cutting down on compute and memory usage and reducing retrieval latency? Vector quantization is how. It helps you store embeddings more compactly, reduce retrieval times, and keep costs under control. Vector quantization offers a powerful solution to scalability, latency, and resource utilization challenges by compressing high-dimensional embeddings into compact representations while preserving their essential characteristics. This technique can dramatically reduce memory requirements and accelerate similarity computations without compromising retrieval accuracy. What is vector quantization? Vector quantization is a compression technique widely applied in digital signal processing and machine learning. Its core idea is to represent numerical data using fewer bits, reducing storage requirements without entirely sacrificing the data’s informative value. In the context of AI workloads, quantization commonly involves converting embeddings—originally stored as 32-bit floating-point values—into formats like 8-bit integers. By doing so, you can substantially decrease memory and storage consumption while maintaining a level of precision suitable for similarity search tasks. An important point to note is that the quantization mechanism is especially suitable for use cases that involve over 1 million vector embeddings, such as RAG applications, semantic search, or recommendation systems that require tight control of operational costs without a compromise on retrieval accuracy. Smaller datasets with fewer than 1 million embeddings might not see significant gains from quantization procedures. For smaller datasets, the overhead of implementing quantization might outweigh its benefits. Understanding vector quantization Vector quantization operates by mapping high-dimensional vectors to a discrete set of prototype vectors or converting them to lower-precision formats. There are three main approaches: Scalar quantization: Converts individual 32-bit floating-point values to 8-bit integers, reducing memory usage of vector values by 75% while maintaining reasonable precision. Product quantization: Compresses entire vectors at once by mapping them to a codebook of representative vectors, offering better compression than scalar quantization at the cost of more complex encoding/decoding. Binary quantization: Transforms vectors into binary (0/1) representations, achieving maximum compression but with more significant information loss. A vector database that applies these compression techniques must effectively manage multiple data structures: Hierarchical navigable small world (HNSW) graph for navigable search Full-fidelity vectors (32-bit float embeddings) Quantized vectors (int8 or binary) When quantization is defined in the vector index, the system builds quantized vectors and constructs the HNSW graph from these compressed vectors. Both structures are placed in memory for efficient search operations, significantly reducing the RAM footprint compared to storing full-fidelity vectors alone. The table below illustrates how different quantization mechanisms impact memory usage and disk consumption. This example focuses on HNSW indexes storing 30 GB of original float32 embeddings alongside a 0.1 GB HNSW graph structure. Our RAM usage estimates include a 10% overhead factor (1.1 multiplier) to account for JVM memory requirements with indexes loaded into page cache, reflecting typical production deployment conditions. Actual overhead may vary based on specific configurations. Here are key attributes to consider based on the table below: Estimated RAM usage: Combines HNSW graph size with either full or quantized vectors, plus a small overhead factor (1.1 for index overhead). Disk usage: Includes storage for full-fidelity vectors, HNSW graph, and quantized vectors when applicable. Notice that while enabling quantization increases total disk usage —because you still store full-fidelity vectors for exact nearest neighbor queries in both cases and rescoring in the case of binary quantization—it dramatically decreases RAM requirements and speeds up initial retrieval . MongoDB Atlas Vector Search offers powerful scaling capabilities through its automatic quantization system . As illustrated in Figure 1 below, MongoDB Atlas supports multiple vector search indexes with varying precision levels: Float32 for maximum accuracy, Scalar Quantized (int8) for balanced performance with 3.75× RAM reduction, and Binary Quantized (1-bit) for maximum speed with 24× RAM reduction. The quantization variety provided by MongoDB Atlas allows users to optimize their vector search workloads based on specific requirements. For collections exceeding 1M vectors, Atlas automatically applies the appropriate quantization mechanism, with binary quantization particularly effective when combined with Float32 rescoring for final refinement. Figure 1: MongoDB Atlas Vector Search Architecture with Automatic Quantization Data flow through embedding generation, storage, and tiered vector indexing with binary rescoring. Binary quantization with rescoring A particularly effective strategy is to combine binary quantization with a rescoring step using full-fidelity vectors. This approach offers the best of both worlds: extremely fast lookups thanks to binary data formats, plus more precise final rankings from higher-fidelity embeddings. Initial retrieval (Binary) Embeddings are stored as binary to minimize memory usage and accelerate the approximate nearest neighbor (ANN) search. Hamming distance (via XOR + population count) is used, which is computationally faster than Euclidean or cosine similarity on floats. Rescoring The top candidate results from the binary pass are re-evaluated using their float or int8 vectors to refine the ranking. This step mitigates the loss of detail in binary vectors, balancing result accuracy with the speed of the initial retrieval. By pairing binary vectors for rapid recall with full-fidelity embeddings for final refinement, you can keep your system highly performant and maintain strong relevance. The need for quantization-aware models Not all embedding models perform equally well under quantization. Models need to be specifically trained with quantization in mind to maintain their effectiveness when compressed. Some models—especially those trained purely for high-precision scenarios—suffer significant accuracy drops when their embeddings are represented with fewer bits. Quantization-aware training (QAT) involves: Simulating quantization effects during the training process Adjusting model weights to minimize information loss Ensuring robust performance across different precision levels This is particularly important for production applications where maintaining high accuracy is crucial. Embedding models like those from Voyage AI— which recently joined MongoDB —are specifically designed with quantization awareness, making them more suitable for scaled deployments. These models preserve more of their essential feature information even under aggressive compression. Voyage AI provides a suite of embedding models specifically designed with QAT in mind, ensuring minimal loss in semantic quality when shifting to 8-bit integer or even binary representations. Figure 2: Embedding model performance comparing retrieval quality (NDCG@10) versus storage costs . Voyage AI models (green) maintain superior retrieval quality even with binary quantization (triangles) and int8 compression (squares), achieving up to 100x storage efficiency compared to standard float embeddings (circles) . The graph above shows several important patterns that demonstrate why quantization-aware training (QAT) is crucial for maintaining performance under aggressive compression. The Voyage AI family of models (shown in green) demonstrates strong performance in retrieval quality even under extreme compression. The voyage-3-large model demonstrates this dramatically—when using int8 precision at 1024 dimensions, it performs nearly identically to its float precision, 2048-dimensional counterpart, showing only a minimal 0.31% quality reduction despite using 8 times less storage. This showcases how models specifically designed with quantization in mind can preserve their semantic understanding even under substantial compression. Even more impressive is how QAT models maintain their edge over larger, uncompressed models. The voyage-3-large model with int8 precision and 1024 dimensions outperforms OpenAI-v3-large (using float precision and 3072 dimensions) by 9.44% while requiring 12 times less storage. This performance gap highlights that raw model size and dimension count aren't the decisive factors —it's the intelligent design for quantization that matters. The cost implications become truly striking when we examine binary quantization. Using voyage-3-large with 512-dimensional binary embeddings, we still achieve better retrieval quality than OpenAI-v3-large with its full 3072-dimensional float embeddings while using 200 times less storage. To put this in practical terms: what would have cost $20,000 in monthly storage can be reduced to just $100 while actually improving performance. In contrast, models not specifically trained for quantization, such as OpenAI's v3-small (shown in gray), show a more dramatic drop in retrieval quality as compression increases. While these models perform well in their full floating-point representation (at 1x storage cost), their effectiveness deteriorates more sharply when quantized, especially with binary quantization. For production applications where both accuracy and efficiency are crucial, choosing a model that has undergone quantization-aware training can make the difference between a system that degrades under compression and one that maintains its effectiveness while dramatically reducing resource requirements. Read more on the Voyage AI blog . Impact: Memory, retrieval latency, and cost Vector quantization addresses the three core challenges of large-scale AI workloads—memory, retrieval latency, and cost—by compressing full-precision embeddings into more compact representations. Below is a breakdown of how quantization drives efficiency in each area. Figure 3: Quantization Performance Metrics: Memory Savings with Minimal Accuracy Trade-offs Comparison of scalar vs. binary quantization showing RAM reduction (75%/96%), query accuracy retention (99%/95%), and performance gains (>100%) for vector search operations Memory and storage optimization Quantization techniques dramatically reduce compute resource requirements while maintaining search accuracy for vector embeddings at scale. Lower RAM footprint Storage in RAM is often the primary bottleneck for vector search systems Embeddings stored as 8-bit integers or binary reduce overall memory usage, allowing significantly more vectors to remain in memory. This compression directly shrinks vector indexes (e.g., HNSW), leading to faster lookups and fewer disk I/O operations. Reduced disk usage in collection with binData binData (binary) formats can cut raw storage needs by up to 66%. Some disk overhead may remain when storing both quantized and original vectors, but the performance benefits justify this tradeoff. Practical gains 3.75× reduction in RAM usage with scalar (int8) quantization Up to 24× reduction with binary quantization, especially when combined with rescoring to preserve accuracy. Significantly more efficient vector indexes, enabling large-scale deployments without prohibitive hardware upgrades. Retrieval latency Quantization methods leverage CPU cache optimizations and efficient distance calculations to accelerate vector search operations beyond what's possible with standard float32 embeddings. Faster similarity computations Smaller data types are more CPU-cache-friendly, which speeds up distance calculations. Binary quantization uses Hamming distance (XOR + popcount), yielding dramatically faster top-k candidate retrieval. Improved throughput With reduced memory overhead, the system can handle more concurrent queries at lower latencies. In internal benchmarks, query performance for large-scale retrievals improved by up to 80% when adopting quantized vectors. Cost efficiency Vector quantization provides substantial infrastructure savings by reducing memory and computation requirements while maintaining retrieval quality through compression and rescoring techniques. Lower infrastructure costs Smaller vectors consume fewer hardware resources, enabling deployments on less expensive instances or tiers. Reduced CPU/GPU time per query allows resource reallocation to other critical parts of the application. Better scalability As data volumes grow, memory and compute requirements don’t escalate as sharply. Quantization-aware training (QAT) models, such as those from Voyage AI, help maintain accuracy while reaping cost savings at scale. By compressing vectors into int8 or binary formats, you tackle memory constraints, accelerate lookups, and curb infrastructure expenses—making vector quantization an indispensable strategy for high-volume AI applications. MongoDB Atlas: Built for Changing Workloads with Automatic Vector Quantization The good news for developers is that MongoDB Atlas supports “automatic scalar” and “automatic binary quantization” in index definitions, reducing the need for external scripts or manual data preprocessing. By quantizing at index build time and query time, organizations can run large-scale vector workloads on smaller, more cost-effective clusters. A common question most developers ask is when to use quantization. Quantization becomes most valuable once you reach substantial data volumes—on the order of a million or more embeddings. At this scale, memory and compute demands can skyrocket, making reduced memory footprints and faster retrieval speeds essential. Examples of cases that call for quantization include: High-volume scenarios: Datasets with millions of vector embeddings where you must tightly control memory and disk usage. Real-time responses: Systems needing low-latency queries under high user concurrency. High query throughput: Environments with numerous concurrent requests demanding both speed and cost-efficiency. For smaller datasets (under 1 million vectors), the added complexity of quantization may not justify the benefits. However, for large-scale deployments, it becomes a critical optimization that can dramatically improve both performance and cost-effectiveness. Now that we have established a strong foundation on the advantages of quantization—specifically the benefits of binary quantization with rescoring— feel free to refer to the MongoDB documentation to learn more about implementing vector quantization. You can also learn more about Voyage AI’s state-of-the-art embedding models on our product page .
Redefining the Database for AI: Why MongoDB Acquired Voyage AI
This post is also available in: Deutsch , Français , Español , Português , Italiano , 한국어 , 简体中文 . AI is reshaping industries, redefining customer experiences, and transforming how businesses innovate, operate, and compete. While much of the focus is on frontier models, a fundamental challenge lies in data—how it is stored, retrieved, and made useful for AI applications. The democratization of AI-powered software depends on building on top of the right abstractions, yet today, creating useful, real-time AI applications at scale is not feasible for most organizations. The challenge isn’t just complexity—it’s trust. AI models are probabilistic, meaning their outputs aren’t deterministic and predictable. This is easily evident in the hallucination problem in chatbots today, and becomes even more critical with the rise of agents, where AI systems make autonomous decisions. Development teams need the ability to control, shape, and ground generated outputs to align with their objectives and ensure accuracy. AI-powered search and retrieval is a powerful tool that extracts relevant contextual data from specific sources, augmenting AI models to generate reliable and accurate responses or take responsible and safe actions, as seen in the prominent retrieval augmented generation (RAG) approach. At the core of AI-powered retrieval are embedding generation and reranking—two key AI components that capture the semantic meaning of data and assess the relevance of queries and results. We believe embedding generation and reranking, as well as AI-powered search, belong in the database layer, simplifying the stack and creating a more reliable foundation for AI applications. By bringing more intelligence into the database, we help businesses mitigate hallucinations, improve trustworthiness, and unlock AI’s full potential at scale. The most impactful applications require a flexible, intelligent, and scalable data foundation. That’s why we’re excited to announce the acquisition of Voyage AI , a leader in embedding and reranking models that dramatically improve accuracy through AI-powered search and retrieval. This move isn’t just about adding AI capabilities— it’s about redefining the database for the AI era . Why this matters: The future of AI is built on better relevance and accuracy in data AI is probabilistic—it’s not built like traditional software with pre-defined rules and logic. Instead, it generates responses or takes action based on how the AI model is trained and what data is retrieved. However, due to the probabilistic nature of the technology, AI can hallucinate. Hallucinations are a direct consequence of poor or imprecise retrieval—when AI lacks access to the right data, it generates plausible but incorrect information. This is a critical barrier to AI adoption, especially in enterprises and for mission-critical use cases where accuracy is non-negotiable. This makes retrieving the most relevant data essential for AI applications to deliver high-quality, contextually accurate results. Today, developers rely on a patchwork of separate components to build AI-powered applications. Sub-optimal choices of these components, such as embedding models, can yield low-relevancy data retrieval and low-quality generated outputs. This fragmented approach is complex, costly, inefficient, and cumbersome for developers. With Voyage AI, MongoDB solves this challenge by making AI-powered search and retrieval native to the database. Instead of implementing workarounds or managing separate systems, developers can generate high-quality embeddings from real-time operational data, store vectors, perform semantic search, and refine results—all within MongoDB. This eliminates complexity and delivers higher accuracy, lower latency, and a streamlined developer experience. What Voyage AI brings to MongoDB Voyage AI has built a world-class AI research team with roots at Stanford, MIT, UC Berkeley, and Princeton and has rapidly become a leader in high-precision AI retrieval. Their technology is already trusted by some of the most advanced AI startups, including Anthropic, LangChain, Harvey, and Replit. Notably, Voyage AI’s embedding models are the highest-rated zero-shot models in the Hugging Face community. Voyage AI’s models are designed to increase the quality of generated output by: Enhancing vector search by creating embeddings that better capture meaning across text, images, PDFs, and structured data. Improving retrieval accuracy through advanced reranking models that refine search results for AI-powered applications. Enabling domain-specific AI with fine-tuned models optimized for different industries such as financial services, healthcare, and law, and use cases such as code generation. By integrating Voyage AI’s retrieval capabilities into MongoDB, we’re helping organizations more easily build AI applications with greater accuracy and reliability—without unnecessary complexity. How Voyage AI will be integrated into MongoDB We are integrating Voyage AI with MongoDB in three phases. In the first phase, Voyage AI’s text embedding, multi-modal embedding, and reranking models will remain widely available through Voyage AI’s current APIs and via the AWS and Azure Marketplaces—ensuring developers can continue to use their best-in-class embedding and reranking capabilities. We will also invest in the scalability and enterprise readiness of the platform to support the increased adoption of Voyage AI’s models. Next, we will seamlessly embed Voyage AI’s capabilities into MongoDB Atlas , starting with an auto-embedding service for Vector Search, which will handle embedding generation automatically. Native reranking will follow, allowing developers to boost retrieval accuracy instantly. We also plan to expand domain-specific AI capabilities to better support different industries (e.g., financial services, legal, etc.) or use cases (e.g., code generation). Finally, we will advance AI-powered retrieval with enhanced multi-modal capabilities, enabling seamless retrieval and ranking of text, images, and video. We also plan to introduce instruction-tuned models, allowing developers to refine search behavior using simple prompts instead of complex fine-tuning. This will be complemented by embedding lifecycle management in MongoDB Atlas, ensuring continuous updates and real-time optimization for AI applications. What this means for developers and businesses AI-powered applications need more than a database that just stores, processes, and persists data—they need a database that actively improves retrieval accuracy, scales seamlessly, and eliminates operational friction. With Voyage AI, MongoDB redefines what’s required for a database to underpin mission-critical AI-powered applications. Developers will no longer need to manage external embedding APIs, standalone vector stores, or complex search pipelines. AI retrieval will be built into the database itself, making semantic search, vector retrieval, and ranking as seamless as traditional queries. For businesses, this translates to faster time-to-value and greater confidence in scaling AI applications. By delivering high-quality results at scale, enterprises can seamlessly integrate AI into their most critical use cases, ensuring reliability, performance, and real-world impact. Looking ahead: What comes next This is just the beginning. Our vision is to make MongoDB the most powerful and intuitive database for modern, AI-driven applications. Voyage AI’s models will soon be natively available in MongoDB Atlas. We will continue evolving MongoDB’s AI retrieval capabilities, making it smarter, more adaptable, and capable of handling a wider range of data types and use cases. Stay tuned for more details on how you can start using Voyage AI’s capabilities in MongoDB. To learn more about how MongoDB and Voyage AI are powering state-of-the-art AI search and retrieval for building, scaling, and deploying intelligent applications, visit our product page .
Multi-Agent Collaboration for Manufacturing Operations Optimization
While there are some naysayers across the media landscape who doubt the potential impact of AI innovations, for those of us immersed in implementing AI on a daily basis, there’s wide agreement that its potential is huge and world-altering. It’s now generally accepted that Large Language Models (LLMs) will eventually be able to perform tasks as well—if not better—than a human. And the size of the potential AI market is truly staggering. Bain’s AI analysis estimates that the total addressable market (TAM) for AI and gen AI-related hardware and software will grow between 40% and 55% annually, reaching between $780 billion and $990 billion by 2027. This growth is especially relevant to industries like manufacturing, where generative AI can be applied across the value chain. From inventory categorization to product risk assessments, knowledge management, and predictive maintenance strategy generation, AI's potential to optimize manufacturing operations cannot be overstated. But in order to realize the transformative economic potential of AI, applications powered by LLMs need to evolve beyond chatbots that leverage retrieval-augmented generation (RAG). Truly transformative AI-powered applications need to be objective-driven, not just responding to user queries but also taking action on behalf of the user. This is crucial in complex manufacturing processes. In other words, they need to act like agents. Agentic systems, or compound AI systems, are currently emerging as the next frontier of generative AI applications. These systems consist of a single or multiple AI agents that collaborate with each other and use tools to provide value. An AI agent is a computational entity containing short- and long-term memory, which enables it to provide context to an LLM. It also has access to tools, such as web search and function calling, that enable it to act upon the response from an LLM or provide additional information to the LLM. Figure 1. Basic components of an agentic system. An agentic system can have more than one AI agent. In most cases, AI agents may be required to interact with other agents within the same system or external systems., They’re expected to engage with humans for feedback or review of outputs from execution steps. AI agents can also comprehend the context of outputs from other agents and humans, and change their course of action and next steps. For example, agents can monitor and optimize various facets of manufacturing operations simultaneously, such as supply chain logistics and production line efficiency. There are certain benefits of having a multi-agent collaboration system instead of having one single agent. You can have each agent customized to do one thing and do it well. For example, one agent can create meeting minutes while another agent writes follow-up emails. It can also be implemented on predictive maintenance, with one agent analyzing machine data to find mechanical issues before they occur while another optimizes resource allocation, ensuring materials and labor are utilized efficiently. You can also provision dedicated resources and tools for different agents. For example, one agent uses a model to analyze and transcribe videos while the other uses models for natural language processing (NLP) and answering questions about the video. Figure 2. Multi-agent collaboration system. MongoDB can act as the memory provider for an agentic system. Conversation history alongside vector embeddings can be stored in MongoDB leveraging the flexible document model. Atlas Vector Search can be used to run semantic search on stored vector embeddings, and our sharding capabilities allow for horizontal scaling without compromising on performance. Our clients across industries have been leveraging MongoDB Atlas for their generative AI use cases , including agentic AI use cases such as Questflow , which is transforming work by using multi-agent AI to handle repetitive tasks in strategic roles. Supported by MiraclePlus and MongoDB Atlas, it enables startups to automate workflows efficiently. As it expands to larger enterprises, it aims to boost AI collaboration and streamline task automation, paving the way for seamless human-AI integration. The concept of a multi-agent collaboration system is new, and it can be challenging for manufacturing organizations to identify the right use case to apply this cutting-edge technology. Below, we propose a use case where three agents collaborate with each other to optimize the performance of a machine. Multi-agent collaboration use case in manufacturing In manufacturing operations, leveraging multi-agent collaboration for predictive maintenance can significantly boost operational efficiency. For instance, consider a production environment where three distinct agents—predictive maintenance, process optimization, and quality assurance—collaborate in real-time to refine machine operations and maintain the factory at peak performance. In Figure 3, the predictive maintenance agent is focused on machinery maintenance. Its main tasks are to monitor equipment health by analyzing sensor data generated from the machines. It predicts machine failures and recommends maintenance actions to extend machinery lifespan and prevent downtime as much as possible. Figure 3. A multi-agent system for production optimization. The process optimization agent is designed to enhance production efficiency. It analyzes production parameters to identify inefficiencies and bottlenecks, and it optimizes said parameters by adjusting them (speed, vibration, etc.) to maintain product quality and production efficiency. This agent also incorporates feedback from the other two agents while making decisions on what production parameter to tune. For instance, the predictive maintenance agent can flag an anomaly in a milling machine temperature sensor reading; for example, if temperature values are going up, the process optimization agent can review the cutting speed parameter for adjustment. The quality assurance agent is responsible for evaluating product quality. It analyzes optimized production parameters and checks how those parameters can affect the quality of the product being fabricated. It also provides feedback for the other two agents. The three agents constantly exchange feedback with each other, and this feedback is also stored in the MongoDB Atlas database as agent short-term memory. In contrast, vector embeddings and sensor data are persisted as long-term memory. MongoDB is an ideal memory provider for agentic AI use case development thanks to its flexible document model, extensive security and data governance features, and horizontal scalability. All three agents have access to a "search_documents" tool, which leverages Atlas Vector Search to query vector embeddings of machine repair manuals and old maintenance work orders. The predictive maintenance agent leverages this tool to figure out additional insights while performing machine root cause diagnostics. Set up the use case shown in this article using our repo . To learn more about MongoDB’s role in the manufacturing industry, please visit our manufacturing and automotive webpage . To learn more about AI agents, visit our Demystifying AI Agents guide .