MongoDB Atlas CLI: Full API Coverage and Faster Updates

Jeroen Vervaeke
June 24, 2025

We’re thrilled to announce that starting today, you can access every feature in the MongoDB Atlas Administration API from the MongoDB Atlas CLI. This significant enhancement means that you’ll also get new features quicker than ever, within just days of their launch. No more hoping for feature support or switching between interfaces. If it’s in the MongoDB Atlas Admin API, it’s in your Atlas CLI.

Full parity with the API

Until now, you had to wait for the Atlas CLI team to manually implement support for new MongoDB Atlas Administration API endpoints. Those days are over. Every MongoDB Atlas capability, whether it launched today or has been around for years, will now automatically become available from your command line.

With the new atlas api subcommands, you get:

  • Full feature parity with the MongoDB Atlas Administration API.

  • Quicker access to future MongoDB Atlas Administration API features.

  • A unified, predictable command structure that makes automation easy.

  • Simplified handling of long-running operations with the ⁠--watch flag, eliminating the need for complex polling logic.

  • The ability to pin a desired API version, ensuring your scripts remain reliable even if you update the CLI.

The reason we built this

We prioritized this feature to ensure our users could use all capabilities exposed by the MongoDB Atlas Administration API through the Atlas CLI without delays. This isn’t about adding new capabilities to MongoDB Atlas but about making existing API functionality accessible through the Atlas CLI, eliminating the previous gap in functionality.

Our goal was simple: If a feature exists in the MongoDB Atlas Administration API, you should be able to access it through the CLI immediately, not weeks or months later. Now you can.

The CLI simplifies API interactions

With the arrival of the complete coverage of the MongoDB Atlas Administration API in the MongoDB Atlas CLI, users no longer need to implement workarounds in order to make use of these endpoints. For example:

  • Authentication: The CLI makes API interactions simpler by automatically handling authentication. This means you don’t have to manage tokens or credentials on your own when making requests to endpoints that need authentication.

  • Monitoring long-running operations: The CLI offers a powerful--watch flag for all API subcommands. This feature automatically monitors long-running operations until they are completed, eliminating the need to implement polling loops manually. Without this flag, you would have to repeatedly check the operation status by directly calling the API.

Let’s take a look at an example of how the --watch flag simplifies waiting for long-running operations.

atlas api clusters createCluster --file clusterspec.json --watch

This command creates a cluster and waits until it’s fully provisioned before returning, eliminating the need for complex polling logic in your scripts.

Practical applications

The atlas api subcommands enable powerful workflows that were previously unavailable in the CLI:

  • Cluster outage simulation: Simulate regional outage scenarios directly through the CLI. You can now script tests that simulate an entire cloud provider’s region being down, helping ensure your applications remain resilient during actual outages.

  • Invoice investigation: Generate custom reports and retrieve billing information programmatically. Need to pull invoice data for your finance team? That’s now a simple CLI command away.

  • Access tracking: Monitor and manage user access patterns across your MongoDB Atlas resources, enhancing your security posture without leaving the command line.

These are just a few of the features now available through the new atlas api subcommands. Visit our documentation to explore the full range of available commands.

Robust and fully documented API subcommands

All atlas api subcommands are auto-generated from our OpenAPI specification, ensuring they stay up to date with the latest Atlas Administration API features. Additionally, these subcommands are versioned, which ensures that your scripts won’t break when the API updates —a critical feature for reliable automation.

For detailed information on syntax and usage, please refer to our comprehensive documentation.

Status: In public preview and ready for your feedback

The introduction of atlas api subcommands represents a significant advancement in making MongoDB Atlas more accessible and automatable. By bringing the full power of the MongoDB Atlas Administration API to the command line, we’re enabling anyone who automates their MongoDB Atlas cloud to work more efficiently.

Whether you’re managing infrastructure, implementing testing protocols, or generating reports, these new capabilities can transform your MongoDB Atlas experience—all without leaving the command line.

As this feature is currently in public preview, we’re actively seeking your input.

Here is how to get started:

Your feedback helps us understand how you’re using these capabilities and what improvements would make them even more valuable to your workflows.

Learn more about the MongoDB Atlas CLI through our documentation.

← Previous

MongoDB and deepset Pave the Way for Effortless AI App Creation

Building robust AI-powered applications has often been a complex, resource-intensive process. It typically demands deep technical and domain expertise, significant development effort, and a long time to value. For IT decision-makers, the goal is clear: enable AI innovation to achieve real business outcomes without compromising scalability, flexibility, or performance, and without creating bottlenecks for development teams serving business teams and customers. Solutions from deepset and MongoDB empower organizations to overcome these challenges, enabling faster development, unlocking AI's potential, and ensuring the scalability and resilience required by modern businesses. Breaking barriers in AI development: The real-time data challenge For many industries, real-time data access is critical to unlocking insights and delivering exceptional customer experiences. AI-driven applications rely on seamless retrieval and processing of structured and unstructured data to fuel smarter decision-making, automate workflows, and improve user interactions. For example, in customer service platforms, instant access to relevant data ensures fast and accurate responses to user queries, improving satisfaction and efficiency. And healthcare applications require immediate access to patient records to enable personalized treatment plans that enhance patient outcomes. Similarly, financial systems rely on real-time analysis of market trends and borrower profiles to make smarter investment and credit decisions to stay competitive in dynamic environments. However, businesses often face challenges when scaling AI applications. These challenges include inconsistent data retrieval, where organizations struggle to efficiently query and access data across vast pools of information. Another challenge is complex query resolution, which involves interpreting multi-layered queries to retrieve the most relevant insights and provide smart recommendations. Data security concerns also pose obstacles, as businesses must ensure sensitive information remains protected while maintaining compliance with regulatory standards. Lastly, AI production-readiness is critical, requiring organizations to ensure their AI applications are properly configured and thoroughly tested to support mission-critical decisions and workflows with accuracy, speed, and adaptability to rapid changes in the AI ecosystem or world events. Addressing these challenges is vital for businesses looking to unlock the full potential of AI-powered innovations and maintain a competitive edge. Transformative solution: Deepset RAG expertise meets MongoDB Atlas Vector Search We’re excited to announce a new partnership between deepset and MongoDB. By integrating deepset’s expertise in retrieval-augmented generation (RAG) and intelligent agents with MongoDB Atlas, developers can now more easily build advanced AI-powered applications that deliver fast, accurate insights from large and complex datasets. We're thrilled to partner with MongoDB and build out an integrated end-to-end GenAI solution to speed up the time to value of customers' AI efforts and help solve their complex use cases to deliver key business outcomes. Mark Ghannam, Head of Partnerships, deepset What sets deepset apart is its product and documentation production-readiness, flexibility for solving complex use cases, and its library of ready-to-use templates, which allow businesses to get started fast to quickly deploy common RAG and agent functionalities, reducing the time and effort required for development. For teams needing customization, Haystack provides a modular, object-oriented design that supports drag-and-drop components , utilizing both standard integrations and custom components . This makes it highly accessible, enabling developers to configure workflows according to their specific application needs, without requiring extensive coding knowledge. On top of Haystack, deepset’s AI Platform makes the prototype to production process of building AI applications even faster and more efficient. It extends Haystack’s building block approach to AI application development, with a visual design interface, qualitative user testing, side-by-side configuration/large language model (LLM) testing, integrated debugging, and hallucination scoring, in addition to expert service assistance and support. The platform’s Studio Edition is free for developers to try. Through seamless integration with MongoDB Atlas Vector Search , deepset equips developers with the ability to incorporate advanced RAG and agent capabilities into their compound AI applications easily through the processes described, known as LLM orchestration. Key features enable several transformative possibilities across industries. Intelligent chatbots allow businesses to deliver precise and context-aware customer interactions, significantly enhancing call center efficiency. Automated content tagging optimizes and streamlines workflows in content management systems, enabling faster categorization and discovery of information. Tailored educational, research, and media platforms personalize learning materials, research, and media content based on user questions and preferences, improving engagement and effectiveness while adhering to institution and brand guidelines. Industry-specific planning systems and workflow automations simplify complex processes, such as lending due diligence. By leveraging the deepset framework alongside MongoDB Atlas Vector Search, developers gain a powerful toolkit to optimize the performance, scalability, and user experience of their applications. This collaboration provides tangible benefits across industries like customer service, content management, financial services, education, defense, healthcare, media, and law—all while keeping complexity to a minimum. Data security and compliance: A foundational priority As organizations adopt advanced AI technologies, protecting sensitive data is paramount. MongoDB Atlas and deepset offer robust protections to safeguard data integrity. MongoDB and deepset provide industry-standard security measures such as encryption, access controls, and auditing, along with compliance certifications like ISO 27001, SOC 2, and CSA STAR. These measures help ensure that sensitive data is handled with care and that client information remains secure, supporting businesses in meeting their regulatory obligations across different sectors. Incorporating MongoDB Atlas into AI solutions allows enterprises using deepset's RAG and Agent capabilities to confidently manage and protect data, ensuring compliance and reliability while maintaining operational excellence. Shaping the future of AI-powered innovation The partnership between MongoDB and deepset is more than a collaboration—it's a driving force for innovation. By merging cutting-edge language processing capabilities with the robust, scalable infrastructure of MongoDB Atlas, this alliance is empowering organizations to create tomorrow's AI applications, today. Whether it’s intelligent chatbots, personalized platforms, or complex workflow automations, MongoDB and deepset are paving the way for businesses to unlock new levels of efficiency and insight. At the core of this partnership is deepset’s advanced RAG and Agent technology, which enables efficient language processing and precise query resolution—essential components for developing sophisticated AI solutions. Complementing this is MongoDB’s reliable cloud database technology, providing unmatched scalability, fault tolerance, and the ability to effortlessly craft robust applications. The seamless integration of these technologies offers developers a powerful toolkit to create applications that prioritize fast time to value, innovation, and precision. MongoDB’s infrastructure ensures security, reliability, and efficiency, freeing developers to focus their efforts on enhancing application functionality without worrying about foundational stability. Through this strategic alliance, MongoDB and deepset are empowering developers to push the boundaries of intelligent application development. Together, they are delivering solutions that are not only highly responsive and innovative but also expertly balanced across security, reliability, and efficiency—meeting the demands of today’s dynamic markets with confidence. Jumpstart your journey Dive into deepset's comprehensive guide on RAG integration with MongoDB Atlas. Then get started with deepset Studio Edition (free) to start building. Transform your data experience and redefine the way you interact with information today! Learn more about MongoDB and deepset's partnership through our partner ecosystem page .

June 24, 2025
Next →

Matryoshka Embeddings: Smarter Embeddings with Voyage AI

In the realm of AI, embedding models are the bedrock of advanced applications like retrieval augmented generation (RAG), semantic search , and recommendation systems. These models transform unstructured data (text, images, audio) into high-dimensional numerical vectors, allowing us to perform similarity searches and power intelligent features. However, traditional embedding models often generate fixed-size vectors, leading to trade-offs between performance and computational overhead. This post will dive deep into Matryoshka Representation Learning (MRL) , a novel approach that creates flexible, multi-fidelity embeddings. We'll compare and contrast MRL with traditional embeddings and quantization, detailing its unique training process and showcasing how Voyage AI's voyage-3-large and the recently released voyage-3.5 models leverage MRL as well as quantization to deliver unparalleled efficiency with MongoDB Atlas Vector Search . Understanding embedding models At their core, embedding models learn to represent discrete items (words, sentences, documents) as continuous vectors in a multi-dimensional space. The key principle is that items with similar meanings or characteristics are mapped to points that are close to each other in this vector space. This spatial proximity then allows for efficient similarity comparisons using metrics like cosine similarity. For example, in a semantic search application, when a user queries "best vegan restaurants," the embedding model converts this query into a vector. It then compares this vector against a database of pre-computed embeddings for restaurant descriptions. Restaurants whose embeddings are "nearby" the query embedding are deemed relevant and returned to the user. Figure 1. Example embedding model. Image Credit:   Hugging Face Blog Challenges with traditional embeddings Historically, embedding models generate vectors of a fixed size, for example, 768, 1024, or 4096 dimensions. While effective, this fixed-size nature presents challenges: Inflexibility: A model trained for, say, 768-dimensional embeddings, will suffer a significant performance drop if you simply truncate its vectors to a smaller size, like 256 dimensions, without retraining. This means you're locked into a specific dimension size, even if a smaller representation would suffice for certain tasks. High computational load: Higher-dimensional vectors demand more computational resources for storage, transfer, and similarity calculations. In scenarios with large datasets or real-time inference, this can lead to increased latency and operational costs. Information loss on truncation: Without specific training, truncating traditional embeddings inevitably leads to substantial information loss, compromising the quality of downstream tasks. Matryoshka Representation Learning MRL, introduced by researchers from the University of Washington, Google Research, and Harvard University in 2022 , offers an elegant solution to these challenges. Inspired by the Russian nesting dolls, MRL trains a single embedding model such that its full-dimensional output can be truncated to various smaller dimensions while still retaining high semantic quality. The magic lies in how the model is trained to ensure that the initial dimensions of the embedding are the most semantically rich, with subsequent dimensions adding progressively finer-grained information. This means you can train a model to produce, say, a 1024-dimensional embedding. Then, for different use cases or performance requirements, you can simply take the first 256, 512, or any other number of dimensions from that same 1024-dimensional vector. Each truncated vector is still a valid and semantically meaningful representation, just at a different level of detail. Figure 2. Matryoshka embedding model truncating the output. Image Credit:   Hugging Face Blog Understanding MRL with an analogy Imagine a movie. A 2048-dimensional MRL embedding might represent the "Full Movie". Truncating it to: 1024 dimensions: Still provides enough information for a "Movie Trailer." 512 dimensions: Gives a "Plot Summary & Movie Details." 256 dimensions: Captures the "Movie Title & Plot One-liner." This "coarse-to-fine" property ensures that each prefix of the full vector remains semantically rich and usable. You simply keep the first N dimensions from the full vector to truncate it. Figure 3. Visualizing the Matryoshka doll analogy for MRL. The unseen hand: How the loss function shapes embedding quality To truly grasp what makes MRL distinct, we must first understand the pivotal role of the loss function in the training of any embedding model. This mathematical function is the core mechanism that teaches these sophisticated models to understand and represent meaning. During a typical training step, an embedding model processes a batch of input data, producing a set of predicted output vectors. The loss function (“J” in the below diagram) then steps in, comparing these predicted embeddings (“y_pred”) against known "ground truth" or expected target values (“y”). It quantifies the discrepancy between what the model predicts and what it should ideally produce, effectively gauging the "error" in its representations. A high loss value signifies a significant deviation – a large "penalty" indicating the model is failing to capture the intended relationships (e.g., placing semantically similar items far apart in the vector space). Conversely, a low loss value indicates accurate capture of these relationships, ensuring that similar concepts (like different images of cats) are mapped close together, while dissimilar ones remain distant. Figure 4. Training workflow including the loss function. The iterative training process, guided by an optimizer, continuously adjusts the model's internal weights with the sole aim of minimizing this loss value. This relentless pursuit of a lower loss is precisely how an embedding model learns to generate high-quality, semantically meaningful vectors. MRL training process The key differentiator for MRL lies in its training methodology. Unlike traditional embeddings, where a single loss value is computed for the full vector, MRL training involves: Multiple loss values: Separate loss values are computed for multiple truncated prefixes of the vector (e.g., at 256, 512, 1024, and 2048 dimensions). Loss averaging: These individual losses are averaged (or summed), to calculate a total loss. Incentivized information packing: The model is trained to minimize this total loss. This process penalizes even the smallest prefixes if their loss is high, strongly incentivizing the model to pack the most crucial information into the earliest dimensions of the vector. This results in a model where information is "front-loaded" into early dimensions, ensuring accuracy remains strong even with fewer dimensions, unlike traditional models where accuracy drops significantly upon truncation. Examples of MRL-trained models include voyage-3-large and voyage-3.5 . MRL vs. quantization It's important to differentiate MRL from quantization, another common technique for reducing embedding size. While both aim to make embeddings more efficient, their approaches and benefits differ fundamentally. Quantization techniques compress existing high-dimensional embeddings into a more compact form, by reducing the precision of the numerical values (e.g., from float32 to int8). The following table describes the precise differences between MRL and Quantization. table, th, td { border: 1px solid black; border-collapse: collapse; } th, td { padding: 5px; } Aspect MRL Quantization Goal Reduce embedding dimensionality (e.g., 256 out of 2048 dims) Reduce embedding precision (e.g., instead of using fp32, using int8/binary embeddings) Output Type Float32 vectors of varying lengths Fixed-length vectors with lower bit representations Training Awareness Uses multi-loss training across dimensions Often uses quantization-aware training (QAT) Use Case Trade-off accuracy vs compute/memory at inference Minimize storage and accelerate vector math operations Example (Voyage AI) voyage-3-large @ 512-dim-fp32 voyage-3-large @ 2048-dim-int8 Flexibility and efficiency with MRL The core benefit of MRL is its unparalleled flexibility and efficiency. Instead of being locked into a single, large vector size, you can: Choose what you need: Generate a full 2048-dimensional vector and then slice it to 256, 512, or 1024 dimensions based on your specific needs. One vector, multiple fidelities: A single embedding provides multiple levels of detail and accuracy. Lower compute, bandwidth, and storage: By using smaller vector dimensions, you drastically reduce the computational load for indexing, query processing, and data transfer, as well as the storage footprint in your database. Efficient computation: The embedding is computed once, and then you simply slice it to the desired dimensions, making it highly efficient. Voyage AI, in particular, leverages MRL by default across its models, including voyage-3-large and the latest voyage-3.5, enabling scalable embeddings with one model and multiple dimensions. This allows you to dynamically choose between space/latency and quality at query time, leading to efficient retrieval with minimal accuracy loss. Voyage AI's dual approach: MRL and quantization for ultimate efficiency Voyage AI models maximize efficiency by combining MRL and quantization. MRL enables flexible embeddings by allowing you to select the optimal vector length—for instance, using 512 instead of 2048 dimensions—resulting in significant reductions in size and computational overhead with minimal accuracy loss. Quantization further compresses these vectors by reducing their bit precision, which cuts storage needs and speeds up similarity search operations. This synergy allows you to choose embeddings tailored to your application’s requirements: a voyage-3-large embedding can be used as a compact 512-dimensional floating-point vector (leveraging MRL) or as a full 2048-dimensional 8-bit integer vector (via quantization). The dual approach empowers you to balance accuracy, storage, and performance, ensuring highly efficient, flexible embeddings for your workload. As a result, Voyage AI models deliver faster inferences and help reduce infrastructure costs when powering applications with MongoDB Atlas Vector Search. Head over to the MongoDB AI Learning Hub to learn how to build and deploy AI applications with MongoDB.

July 14, 2025