Boost Connected Car Developments with MongoDB Atlas and AWS

Igor Alekseev (AWS), Humza Akhtar, Rami Pinto Prieto, Raphael Schor, and Anuj Panchal
August 11, 2025

As vehicles continue to evolve from mechanical systems to connected, software-defined platforms, the automotive industry is continuously being reshaped by data. With modern cars generating terabytes of sensor data daily, a key challenge facing the industry is how to extract timely, actionable insight from that data. And a recent survey by McKinsey underscored the degree to which strong connectivity is important to car buyers—close to 40% of US survey respondents indicated that they are willing to switch OEMs over better connectivity options. Though connectivity preferences vary widely by country, autonomous driving and safety features are top of mind for many customers.

In such a landscape, OEMs need to offer new innovative use cases on top of the customer data. For example, one of MongoDB’s large automotive clients is combining car telemetry data with engine noise to perform faster diagnostics and maintenance services. Combining car telemetry data and Internet of Things (IoT) infrastructure with generative AI unlocks enormous potential for the auto manufacturers, from predictive maintenance and remote diagnostics to context-aware driver assistance, smart infotainment, and usage-based insurance models.

Imagine a vehicle that not only warns of a failing battery but also proactively recommends the nearest certified service center with the right parts in stock. A fleet manager might analyze driving behavior across hundreds of trucks to optimize fuel efficiency and reduce accident risks. And with improved data, manufacturers could aggregate warranty and performance data across regions to detect early signs of systemic issue, responding before small defects become expensive recalls. Insurance providers, meanwhile, might use real-time driving profiles to offer policies tailored to individual habits, replacing static risk models with dynamic pricing.

To enable such use cases, organizations require a scalable, flexible and secure data infrastructure. MongoDB Atlas not only offers a flexible document data model but also built-in time series support, high availability, geospatial indexing, and horizontal scalability to handle millions of connected vehicles and associated use cases and services. Combined with AWS services for IoT, edge processing, machine learning, and generative AI, this stack becomes a robust foundation for intelligent mobility.

This blog post explores how enterprises can build such a connected car architecture using MongoDB Atlas, Atlas Vector Search, AWS IoT Greengrass, Amazon Bedrock, and LangChain—as shown in Figure 1. We will convert raw automotive telemetry into real-time business value for drivers, technicians, and fleet managers using an example of car maintenance business workflow.

Figure 1. Connected vehicle data platform architecture with MongoDB Atlas and AWS.
Diagram showing the connected vehicle data platform. The vehicle produces data that is sent through AWS and MongoDB tooling to product information for the user and the technician.

The limitations of traditional maintenance models

Vehicle maintenance still follows two basic patterns: reactive and scheduled. In the reactive model, service is initiated only after a problem has already impacted car performance. At that point, it is too late to avoid costly repairs. Scheduled maintenance is more proactive but is often inefficient, leading to unnecessary servicing that proves costly for the driver and does not reflect actual wear and usage.

The automotive sector needs to shift toward predictive and personalized care, relying on the connected car data that is being collected in real time by OEMs. But achieving this requires a cloud-native data infrastructure that can support continuous ingestion and real-time processing of this data.

From raw sensor data to driving insight

The connected vehicle data journey begins at the edge. Vehicle operational data—from engine RPM and temperature, to battery voltage, tire pressure, and onboard diagnostic codes—can be processed locally on the car using AWS IoT Greengrass, a service from AWS that enables local decision-making even without constant cloud connectivity.

From there, the data flows into AWS IoT Core and is published to Amazon MSK (Managed Streaming for Apache Kafka). Atlas Stream Processing—which ensures scalable, fault-tolerant stream processing—connects to MSK and ingests this data into MongoDB Atlas, where it is stored using a schema modeled on the Vehicle Signal Specification (VSS), a standard developed by the COVESA alliance.

VSS is a widely adopted open data model that helps normalize vehicle signals and improve interoperability, and it provides a hierarchical, standardized format for structuring vehicle data. It defines a semantic tree of signal, such as Vehicle.Speed, Vehicle.Powertrain.Engine.RPM, or Vehicle.Cabin.Door.FrontLeft.IsOpen to ensure consistency and interoperability across makes, models, and applications. This consistency is critical for large-scale data analysis, cross-platform integration, and AI training.

MongoDB, an active member of the COVESA community, is particularly well-suited to implement VSS. Our document-oriented data model allows developers to store deeply nested, flexible JSON structures without enforcing rigid and normalized schemas. This is especially useful when working with evolving vehicle software platforms or optional equipment packages, trim levels, etc., that alter the signal tree. Whether a car has two doors or four, a combustion or an electric drive, MongoDB can seamlessly adapt to its VSS-defined structure without structural rework, saving time and money for the OEMs.

Once vehicle data lands in MongoDB Atlas, a series of event-driven triggers enable real-time reactions. Atlas Triggers can detect when an engine temperature exceeds safe thresholds and immediately invoke an AWS Lambda function to log the incident, notify support teams via Amazon EventBridge, or create a maintenance task in a service management system.

A strong data tiering strategy is important for connected vehicle use cases. For longer-term trend analysis, vehicle data can be exported to Amazon S3 for model training in Amazon SageMaker. These models can forecast component wear, detect behavioral anomalies, or estimate the Remaining Useful Life (RUL) of key systems. Once the model is trained, it can infer directly on the MongoDB data and feed prediction results back into the database, closing the loop.

The alerts and raw telemetry can live inside MongoDB time series collections, which are optimized for high-speed time series data storage and processing. Time series collections also come with Window Functions that enable operations on a specified span of documents or a window in time.

Empowering technicians with AI and vector search

Once an alert is raised, we can use gen AI to enhance customer and technician experience in dealing with and resolving the identified issue. Traditional diagnostic workflows involve sifting through manuals, logs, and systems of record. Now, with Amazon Bedrock and Atlas Vector Search, technicians can simply ask natural-language questions using a chat assistant embedded in a mobile or web application.

Unstructured data such as service manuals, historical record, and technical bulletins are vectorized into arrays of embeddings. These embeddings are indexed and stored in MongoDB Atlas. Once stored and indexed, the technician can query “What is the root cause of the service engine light?” and Atlas Vector Search can search through the vector embeddings and retrieve the most relevant, semantically aligned documents. These results can be fed into large language models exposed by AWS Bedrock to generate the response in a conversational language and tone.

MongoDB’s vector search capability integrates seamlessly with traditional metadata search, combining structured queries (e.g., vehicle ID, timestamp) with semantic matching. This unified approach enhances technician productivity and shortens repair cycles—resulting in positive customer engagement.

To expose this data (and these insights) to different users, we can leverage AWS AppSync as a managed GraphQL interface. Through AppSync, users can query live telemetry, view predicted maintenance needs, or trigger actions like assigning a technician or updating a vehicle’s diagnostic state—ensuring consistency between backend services and user-facing applications.

Business impact across automotive domains

The potential applications of this architecture span the entire automotive value chain.

For example, fleet operators could benefit from predictive service scheduling, improving uptime while reducing costs. Manufacturers would gain insights into failure patterns, enabling them to make data-driven decisions about component design or supplier quality. Dealerships can improve first-time fix rates with AI-guided diagnostics, while insurance companies could implement usage-based models grounded in real driving behavior. Even suppliers and logistics chains could benefit, using aggregated data to anticipate demand and optimize inventory levels.

Smart vehicles, smart connectivity

MongoDB’s high-performance, scalable database—paired with the IoT, AI, and machine learning capabilities of AWS—creates a responsive, resilient connected car platform.

As vehicles grow smarter, so too must the systems that manage their data. MongoDB’s alignment with the VSS standard ensures that automotive data remains interoperable, searchable, and AI-ready. Atlas Vector Search ensures efficient retrieval of context stored in unstructured data, and when paired with AWS services like IoT Greengrass, SageMaker, Bedrock, and AppSync, this architecture allows enterprises to scale to millions of connected vehicles with confidence. For more information on how to model data in MongoDB using VSS specification, check out our other article.

To see these concepts in action, visit our GitHub repository for a hands-on experience and detailed instructions. To learn more about MongoDB’s role in the manufacturing industry, please visit our manufacturing and automotive page.

← Previous

You Don't Always Need Frontier Models to Power Your RAG Architecture

Frontier AI models are driving the widespread adoption of generative AI by demonstrating unprecedented capabilities. However, their deployment often entails significant costs. The strategic partnership between MongoDB and Fireworks.AI addresses these cost implications by offering solutions that optimize performance and resource utilization. This collaboration leverages MongoDB's efficient data management alongside Fireworks.AI's model optimization tools to enhance speed and efficiency while minimizing operational expenses. In the current AI environment, achieving high performance is crucial, but equally important is optimizing the total cost of ownership (TCO). Businesses must focus on the price-performance ratio, ensuring that improvements in speed and efficiency lead to real cost savings. This article will address the following topics: How to build an agentic RAG using a Fireworks AI hosted LLM and MongoDB Atlas for retrieval. Strategies for optimizing retrieval-augmented generation (RAG) applications using MongoDB Atlas and large language models (LLMs) through effective query and response caching. Techniques using the Fireworks AI platform focus on fine-tuning models, accelerating LLM inference, and reducing hardware needs. Steps to fine-tune a pretrained SLM with PEFT techniques using the Fireworks Platform. Readers will gain a practical, in-depth strategy to improve AI performance while lowering costs. This will be demonstrated with examples and performance data. Unlocking efficiency and performance with MongoDB and Fireworks AI MongoDB Atlas is renowned for its flexible schema, efficient indexing, and distributed architecture, allowing organizations to scale their data infrastructure on demand. MongoDB Atlas is a general-purpose database that focuses on highlighting flexibility, AI suitability, and ACID transactions. Users have the flexibility to run their application anywhere but making sure that there are no compromises made in the security aspects of working with it. MongoDB offers a comprehensive, secure, and efficient database solution for modern applications, catering to various technical and strategic needs. Fireworks AI is recognized for its suite of technologies focused on optimizing the performance and efficiency of large language models (LLMs). Their offerings span model optimization tools, a specialized FireOptimizer framework, and innovative attention mechanisms like FireAttention. These solutions aim to enhance inference speeds, reduce operational costs, and improve resource utilization. Furthermore, Fireworks AI provides parameter-efficient fine-tuning methods and adaptive speculative execution to tailor models for specific applications. Their advancements also include optimized processing for long-context tasks and techniques to maximize throughput and cost-effectiveness in model serving. Fireworks also provides model serving functionality for select models that are readily available, also they do provide a platform to host and serve custom implementations of LLM models for customers. Core capabilities: FireOptimizer and FireAttention The FireOptimizer is Fireworks.ai’s adaptation engine for customizing AI model performance in production environments. It automates latency and quality optimization for unique inference workloads. It tailors performance across hardware, model, and software layers using techniques like customizable quantization, fine-tuning, and adaptive caching. Its hallmark feature, adaptive speculative execution, automatically trains workload-specific draft models to parallelize token generation, achieving up to 3x latency improvements compared to generic speculative decoding. This method significantly boosts responsiveness without compromising accuracy by increasing the hit rate. Figure 1. FireOptimizer platform. FireAttention, Fireworks AI's custom-built inference engine, significantly enhances LLM inference speed on GPUs. It achieves this by utilizing a novel micro-precision data format and rewriting key GPU kernels (such as attention and matrix multiplication) from scratch, aligning them with underlying hardware instructions. While FireAttention prioritizes speed, potentially at the cost of initial accuracy, this is mitigated through Quantization-Aware Training (QAT). This approach allows finetuned models to maintain high precision while reducing their memory footprint. Benchmarks demonstrate FireAttention V4's superior performance over SGLang on H200 and TRT-LLM on B200, particularly in MMLU Pro tests. Overall, FireAttention V4 represents a breakthrough in achieving low-latency, high-efficiency LLM inference, especially beneficial for frontier models like DeepSeek R1. Key benefits: Faster inference: FireOptimizer's adaptive speculative execution has demonstrated up to 3x latency improvements in production workloads across various models, ensuring highly responsive applications. Hassle-free optimization: FireOptimizer automates the complexities of optimization, allowing users to concentrate on application development. FireOptimizer FireOptimizer improves batch inference by integrating with MongoDB for efficient model fine-tuning and streamlined deployment. This multi-layered customization is vital for compound AI systems, ensuring consistent model alignment. Available for enterprise on-premise and own-cloud, FireOptimizer enhances traditional inference performance through techniques like adaptive speculative execution, caching, customizable quantization, personalized fine-tuning at scale, and customizable hardware mapping. In this blog post, we’ll explore the benefits of FireOptimizer to perform Parameter-Efficient Fine-Tuning (PEFT) so we can use a small language model(SLM) model to carry out personalized tasks such as RAG for a private dataset. This activity will demonstrate how generative AI can be adopted for general use at scale and critical domains effectively. Survey of fine-tuning strategies for smaller, efficient models Smaller language models present significant opportunities for tailored adaptation while using fewer resources. The ongoing evolution in this field is fueled by increasing demand for deploying optimized LLMs across diverse environments, including cloud platforms, edge devices, and specialized hardware. These fine-tuning approaches can be categorized as follows: Additive parameter-efficient fine-tuning (PEFT): This class of methods augments pre-trained models with new trainable parameters without altering the original weights. Adapters: These involve inserting small, trainable modules, known as adapters, within the pre-trained model's layers. These adapters learn task-specific adjustments, enabling adaptation to new tasks without changing the pre-existing parameters. Soft prompts: These are trainable vector embeddings appended to the input sequence, acting as guiding signals to influence the model's output for a specific task. Prefix tuning: This technique adds a trainable prefix to the input sequence. This prefix learns task-specific information without requiring modifications to the core model architecture. Reparametrization PEFT: This approach reduces the number of trainable parameters by reparameterizing existing model weights using low-rank approximations. Low-Rank Adaptation (LoRA): LoRA approximates weight updates in the attention layers of a pre-trained model using low-rank matrices, significantly decreasing the number of trainable parameters. Quantized LoRA (QLoRA): QLoRA builds upon LoRA by integrating quantization methods, further decreasing memory footprint and computational expenses. Selective fine-tuning: This category focuses on fine-tuning only specific parameters of the pre-trained model, leading to improved computational efficiency. BitFit: This method fine-tunes only the bias terms, or other designated parameters, of the pre-trained model, enhancing computational efficiency. DiffPruning: This technique identifies and removes parameters that have minimal impact on the model's performance, thus reducing the number of trainable parameters. Layer freezing strategies: These strategies involve selectively freezing certain layers of the pre-trained model while fine-tuning others to optimize the adaptation process. Freeze and reconfigure (FAR): FAR involves freezing specific layers of the pre-trained model and fine-tuning the remaining layers to optimize model adaptation. FishMask: This technique uses a mask to selectively freeze or fine-tune layers, optimizing adaptation for specific tasks. Parameter-Efficient Fine-Tuning (PEFT) is a popular technique for adapting small pre-trained models to niche tasks. By adjusting only a small portion of the model's parameters, PEFT prevents overfitting, especially on smaller datasets, and greatly reduces computational and memory demands compared to full fine-tuning. Additionally, PEFT helps mitigate catastrophic forgetting in LLMs. This approach allows for efficient model customization in resource-constrained environments without the need for complete retraining. Leveraging PEFT LoRA techniques in Fireworks AI, combined with the availability of trace data and labeled data, allows for efficient fine-tuning of smaller models. To demonstrate the practical implications of using a small language model (SLM), we will build an agentic RAG application using MongoDB Atlas and illustrate how MongoDB can be used to power semantic search capabilities and also be leveraged as a semantic caching layer. The application serves as a demonstration to follow along with a step-by-step guide to build a simple application that is task-driven by using a Frontier LLM model, such as Llama Maverick, and they fine-tune using data generated out of this setting to fine-tune an SLM that will satisfactorily perform a similar operation while consuming fewer resources. Step-by-Step guide for building an Agentic RAG application with MongoDB Atlas The sample code below demonstrates an end-to-end Agentic Retrieval-Augmented Generation (RAG) workflow using LangChain, MongoDB Atlas Vector Search, and Fireworks LLMs. Below is a summary of the key steps and components: 1. Data loading & preprocessing PDF loading: The EU Act regulations PDF is loaded using PDFLoader . Text splitting: The document is split into manageable chunks using RecursiveCharacterTextSplitter for efficient retrieval and embedding. 2. Embedding & vector store setup Embeddings: Sentence-transformers MPNet model is used to generate vector embeddings for each text chunk. MongoDB Atlas Vector Search : The embeddings and text chunks are stored in MongoDB, and a vector search index is created for similarity search. 3. LLM & caching LLM setup: Meta Llama Maverick is used as the main LLM, with a custom output parser to clean up responses. Semantic cache: MongoDB Atlas Semantic Cache is configured to cache LLM responses and avoid redundant computation./li> 4. Agentic RAG workflow StateGraph Construction: The workflow is modeled as a state machine with the following steps: plan_step: Reformulates the user query for optimal retrieval. retrieve_documents_step: Retrieves relevant documents from the vector store. execute_step: Generates an answer using the LLM and the retrieved context. validate_step: Uses the LLM to validate the relevance of the answer. should_continue: Decides whether to proceed to the execute step or go back to the plan step. Steps to build the Agentic RAG as described above are available in the notebook here . Once built, the graph for your agentic workflow looks as shown in Figure 2. Figure 2. Agentic RAG workflow graph. Running the Agentic RAG Workflow Invoke the workflow with a user query: query = "In the EU act what are the various biometric categorizations of data?" app.invoke({"question": query}, config={"recursion_limit": 30}) Response: Response: In EU legislation, biometric data is categorized and regulated under several key frameworks, primarily focusing on its use, protection, and specific applications. Based on the provided context and relevant regulations: ### 1. **Biometric Verification (Authentication)** - **Definition**: Automated one-to-one comparison of biometric data (e.g., fingerprints, facial recognition) to verify an individual’s identity against previously stored data. - **Purpose**: Authentication (e.g., unlocking devices, accessing services). - **Legal Reference**: Article 3(36) of the cited regulations. ### 2. **Biometric Identification (Matching)** - **Definition**: One-to-many comparison of biometric data against databases to establish identity (e.g., border control, law enforcement). - **Examples**: - **Eurodac** (Regulation 2024/1358): Compares biometric data (e.g., fingerprints) to identify illegally staying third-country nationals or stateless persons. - **Law Enforcement**: Used to locate suspects, victims of trafficking, or missing persons under strict conditions (Article 3(38), 3(39), and provisions like point (i)–(iii)). ### 3. **Special Categories of Personal Data** - **Status**: Biometric data is classified as a "special category" under: - **GDPR (Article 9(1), Regulation (EU) 2016/679)**: Requires enhanced protections due to its sensitivity. - **Directive (EU) 2016/680** and **Regulation (EU) 2018/1725**: Extend these protections to law enforcement and EU institutions. - **Safeguards**: Pseudonymization, strict access controls, confidentiality obligations, and mandatory deletion after retention periods (points (c)–(e) in the context). ### 4. **Operational and Sensitive Data** - **Sensitive Operational Data**: Biometric data used in criminal investigations or counter-terrorism, where disclosure could jeopardize proceedings (Article 3(38)). - **Emotion Recognition Systems**: While not explicitly labeled as biometric, these systems infer emotions/intentions (Article 3(39)) and may intersect with biometric processing if tied to identifiable individuals. ### 5. **Law Enforcement Exceptions** - Biometric data may be processed for: - Preventing terrorist attacks or imminent threats (point (ii)). - Investigating serious crimes (punishable by ≥4 years’ imprisonment) under Annex II (point (iii)). ### Key Requirements: - **Security**: State-of-the-art measures, pseudonymization, and access documentation (point (c)). - **Restrictions**: Prohibition on unauthorized transfers (point (d)). - **Retention**: Deletion after correcting bias or reaching retention limits (point (e)). These categorizations ensure biometric data is used proportionally, with stringent safeguards to protect privacy and fundamental rights under EU law. Validation Score: Score: 0.9 This notebook provides a modular, agentic RAG pipeline that can be adapted for various document retrieval and question-answering tasks using MongoDB and LLMs. Step-by-Step guide for fine-tuning a small language model with Fireworks AI Current challenges with frontier models The large language model used in the preceding example, accounts/fireworks/models/deepseek-r1 , can result in slow application response times due to the significant computational resources required for its billions of parameters. An agentic RAG task involves multiple LLM invocations for steps such as generating retrieval questions, producing answers, and comparing user questions to the generated results. This process involves several LLM queries, extending the total response time to 30-40 seconds , with each query potentially taking 5 or more seconds . Additionally, deploying and scaling LLMs for a large user base can be complex and expensive. To mitigate this issue, the example code demonstrates the use of a semantic cache; however, this only addresses repeated queries to the system. By leveraging small language models (SLMs), enterprises can achieve significant gains in processing speed and cost-efficiency. SLMs require less computational power, making them ideal for resource-constrained devices, while delivering faster response times and lower operational costs. But there is a huge caveat using SLM; they come with several limitations, such as reduced generalization, limited context retention, and lower accuracy on complex tasks compared to larger models. They may struggle with nuanced reasoning, exhibit increased biases, and generate hallucinations due to their constrained training data and fewer parameters. While they are computationally efficient and well-suited for lightweight applications, their ability to adapt across domains remains restricted; for example, a pretrained SLM such as accounts/fireworks/models/deepseek-r1-distill-qwen-1p5b does not produce results which is satisfactory in our agentic RAG setting. It is not able to perform validation scoring or tends to hallucinate, which generatesa response even when context is provided. Adapting a pre-trained Small Language Model (SLM) for specialized applications such as agentic Retrieval-Augmented Generation (RAG) utilizing private knowledge bases offers a cost-effective alternative to frontier models while maintaining similar performance levels. This strategy also provides scalability for numerous clients, ensuring Service Level Agreements (SLAs) are met. Parameter-Efficient Fine-Tuning(PEFT) i.e. QLoRA techniques, including Quantized Low-Rank Adaptation (LoRA), substantially improve efficiency by focusing optimization on a limited set of parameters. This method lowers memory demands and operational expenses. Integrating with MongoDB streamlines data management and supports efficient model fine-tuning workflows. MongoDB's unique value MongoDB is integral, providing seamless data management and real-time integration that improves operational efficiency. By storing trace data as JSON and enabling efficient retrieval and storage, MongoDB adds substantial value to the process of fine-tuning models. MongoDB also doubles up as a caching layer to avoid unnecessarily invoking LLM on repeated requests for the same data. The following steps will go through step-by-step how one can make use of the platform to fine-tune an SLM. Here’s how to leverage this platform and tool: Figure 3. The fine-tuning process explained. To enhance RAG applications, the initial step involves collecting data relevant to the specific task for fine-tuning. MongoDB Atlas, a flexible database, can be utilized to store LLM responses in a cache. For example, in our agentic RAG approach, we can create questions using diverse datasets and store their corresponding answers in MongoDB Atlas. While a powerful LLM might be useful for generating these initial responses or task-specific data during this simulation phase, a smaller scale fine-tuning process requires at least 1000 examples. Subsequently, these generated responses need to be converted into the required format for the Fireworks.ai platform to begin the fine-tuning process. The cache.jsonl file, used later in fine-tuning, can be created by executing the provided code. from pymongo import MongoClient import pandas as pd import json client = MongoClient("<mongodb_atlas_connection_string>") cache_col = client["agenticrag"]["cache"] df = pd.DataFrame.from_records(cache_col.find()) vals = list(zip([{"role": "user", "content": json.loads(text)[0]["kwargs"]["content"]} for text in df.text], [ {"role": "assistant", "content": json.loads(json.loads(text)[0])["kwargs"]["text"]} for text in df.return_val])) messages = [] for val in vals: messages += [{"messages": list(val)}] with open("cache.jsonl", "w") as f: for item in messages: f.write(json.dumps(item) + "\n") Now that we have prepared the dataset and generated our cache.jsonl file, we can fine-tune the pre-trained deepseek-r1-distill-qwen-1p5b model by following the steps below. Prerequisites: Install firectl: Use the command pip install firectl to install the Fireworks command-line tool. Authenticate: Log in to your Fireworks account using firectl login. Prepare Dataset: Ensure your fine-tuning dataset (created during the data generation process) is ready. Steps: 1. Upload dataset: Upload your prepared dataset to the Fireworks platform using the following command, replacing <dataset_name> with your desired name and cache.jsonl with your dataset file: 2. firectl create dataset <dataset_name> cache.jsonl 3. Create fine-tuning job: Initiate a fine-tuning job by specifying the base model, dataset, output model name, LoRA rank, and number of epochs. For example: 4. firectl create sftj --base-model accounts/fireworks/models/deepseek-r1-distill-qwen-1p5b \ 5. --dataset <dataset_name> --output-model ragmodel --lora-rank 8 --epochs 1 6. The output will provide details about the job, including its name, creation time, dataset used, current state, and the name of the output model. 7. Monitor fine-tuning: Track the progress of your fine-tuning job using the Fireworks AI portal. This allows you to ensure the process is running as expected. 8. Deploy fine-tuned model: Once the fine-tuning is complete, deploy the model for inference on the Fireworks platform. This involves two steps: Deploy the base model used for fine-tuning: firectl create deployment accounts/fireworks/models/deepseek-r1-distill-qwen-1p5b --enable-addons --wait Deploy the fine-tuned LoRA adapter: firectl load-lora ragmodel --deployment <deployment_id> 9. Use deployed model: After deployment, the model ID (e.g., models/ragmodel ) can be used to invoke the fine-tuned language model via your preferred LLM framework, leveraging the Fireworks platform's serverless API. Summary Fine-tuning smaller language models (SLMs) for Retrieval Augmented Generation (RAG) using platforms like Fireworks AI offers significant advantages over relying solely on large frontier models. This approach drastically improves response times, reducing latency from around 5 seconds with a large LLM to 2.3 seconds with a fine-tuned SLM , while also substantially decreasing memory and hardware requirements. By leveraging parameter-efficient fine-tuning techniques and integrating with data management solutions like MongoDB, businesses can achieve faster, more cost-effective AI performance for RAG applications, making advanced AI capabilities more accessible and sustainable. Conclusion The collaboration between MongoDB and Fireworks AI offers a powerful synergy for enhancing the efficiency and affordability of Large Language Model (LLM) training and deployment. Fireworks AI's utilization of Parameter-Efficient Fine-Tuning (PEFT) techniques like LoRA and qLoRA significantly curtails the computational resources necessary for fine-tuning LLMs by focusing on low-rank adaptation and quantization. This directly translates to substantial reductions in the costs associated with this crucial process. Complementarily, MongoDB's robust infrastructure, characterized by its distributed architecture, flexible schema, and efficient indexing capabilities, provides the ideal data management foundation. It allows for on-demand scaling of data infrastructure while minimizing storage expenses, thereby contributing to lower capital and operational expenditures. This integration further fosters streamlined workflows between data and AI processes. MongoDB's capacity for real-time data integration ensures that AI models have immediate access to the most current information, thereby improving operational efficiency and the relevance of the models' insights. When combined with Fireworks AI's fine-tuning tools, this creates a cohesive environment where AI models can be continuously updated and refined. Moreover, the partnership simplifies the development of robust Retrieval Augmented Generation (RAG) solutions. MongoDB Atlas offers a scalable platform for storing embeddings, while Fireworks AI provides managed LLM hosting and other essential features. This seamless combination enables the creation of scalable and intelligent systems that significantly enhance user experience through more effective and relevant information retrieval. Organizations adopting this strategy can achieve accelerated AI performance, resource savings, and future-proof solutions—driving innovation and competitive advantage across different sectors. Further reading: Atlas Vector Search : Learn AI and vector search; generate, store, index, and search embeddings in MongoDB Atlas for semantic search. Build hybrid search with Atlas Search and Atlas Vector Search. Use vector search for a RAG chatbot. Manage indexes with Atlas CLI and MongoDB Shell. FireAttention V4 : Enables cost-effective GPU inference and provides industry-leading latency and cost efficiency with FP4. FireOptimizer : Allows users to customize latency and quality for production inference workloads.

August 11, 2025
Next →

Smarter AI Search, Powered by MongoDB Atlas and Pureinsights

We’re excited to announce that the integration of MongoDB Atlas with the Pureinsights Discovery Platform is now generally available—bringing to life a reimagined search experience powered by keyword, vector, and gen AI. What if your search box didn’t just find results, but instead understood intent? That’s exactly what this integration delivers! Beyond search: From matching to meaning Developers rely on MongoDB’s expansive knowledge ecosystem to find answers fast. But even with a rich library of technical blogs, forum threads, and documentation, traditional keyword search often falls short—especially when queries are nuanced, multilingual, or context-driven. That’s where the MongoDB-Pureinsights solution shines. Built on MongoDB Atlas and orchestrated by the Pureinsights Discovery platform, this intelligent search experience starts with the fundamentals: fast, accurate keyword results, powered by MongoDB Atlas Search . But as queries grow more ambiguous—say, “tutorials for AI”—the platform steps up. MongoDB Atlas Vector Search with Voyage AI , available as an embedding and reranking option (now part of MongoDB), goes beyond literal keywords to interpret intent—helping applications deliver smarter, more relevant results. The outcome: smarter, semantically aware responses that feel intuitive and accurate—because they are. What’s more, with generative answers enabled, the platform synthesizes information across MongoDB’s ecosystem (blog content, forums, and technical docs) to deliver clear, contextual answers using state-of-the-art language models. But it's not just pointing you to the right page. Instead, the platform is providing the right answer, with citations, ready to use. It’s like embedding a domain-trained AI assistant directly into your search bar. “As organizations look to move beyond traditional keyword search, they need solutions that combine speed, relevance, and contextual understanding,” said Haim Ribbi, Vice President, Global CSI, VAR & Tech Partner at MongoDB. “MongoDB Atlas provides the foundation for smarter discovery, and this collaboration with Pureinsights shows how easily teams can deliver gen AI-powered search experiences using their existing content.” Built for users everywhere But intelligence alone doesn’t make it transformational. What sets this experience apart is its adaptability. Whether you’re a developer troubleshooting in Berlin or a product owner building in São Paulo, the platform tailors responses to your preferences. Prefer concise summaries or deep technical dives? Want to translate answers in real time? Need responses that reflect your role and context? You’re in control. From tone and length to language and specificity, this is a search that truly understands you—literally and figuratively. Built on MongoDB. Elevated by Voyage AI. Delivered by Pureinsights. At the core of this solution is MongoDB Atlas, which unifies fast, scalable data access to structured content through Atlas Search and Atlas Vector Search. Looking ahead, by integrating with Voyage AI’s industry-leading embedding models, MongoDB Atlas aims to make semantic search and retrieval-augmented generation (RAG) applications even more accurate and reliable. While currently in private preview, this enhancement signals a promising future for developers building intelligent, AI-powered experiences. Pureinsights handles the orchestration layer. Their Discovery Platform ingests and enriches content, blends keyword, vector, and generative search into a seamless UI, and integrates with large language models like GPT-4. The platform supports multilingual capabilities, easy deployment, and enterprise-grade scalability out of the box. While generative answers are powered by integrated large language models (LLMs) and may vary by deployment, the solution is enterprise-ready, cloud-native, and built to scale. Bringing intelligent discovery to your own data Watch the demo video to see AI-powered search in action across 4,000+ pages of MongoDB content—from community forums and blog posts to technical documentation. While the demo features MongoDB’s content, the solution is built to adapt. You can bring the same AI-powered experience to your internal knowledge base, customer support portal, or developer hub—no need to build from scratch. Visit our partner page to learn more about MongoDB and Pureinsights and how we’re helping enterprises build smarter, AI-powered search experiences. Apply for a free gen AI demo using your enterprise content.

October 1, 2025