Evans Future-Ready Data Foundation with MongoDB

Principles for designing scalable and AI-ready data architectures
1. Auflage 2026
ISBN: 978-1-80760-716-6
Verlag: Packt Publishing
Format: EPUB
Kopierschutz: 0 - No protection

E-Book, Englisch, 98 Seiten

AI success depends on more than models, it requires a data foundation built for scale, intelligence, and trust. In Future-Ready Data Foundation with MongoDB, you'll discover the principles behind building AI-ready data architectures that support modern applications, retrieval-augmented generation (RAG), and AI agents. This concise guide explores the critical role of data in AI modernization and shows how MongoDB helps organizations create a unified foundation for innovation.
Through practical architectural insights, you'll learn how retrieval strategies influence the quality and reliability of AI outcomes and how MongoDB's document model and vector search capabilities support intelligent data access. You'll also explore the scalability, performance, and operational patterns required to keep AI systems running efficiently, including replication, sharding, workload isolation, and search optimization. Finally, you'll examine the governance, observability, security, and compliance considerations that help organizations deploy AI responsibly and at scale.
By the end of this book, you'll be equipped to evaluate, design, and optimize data foundations that support the next generation of AI-powered applications.

Autoren/Hrsg.

Evans, Sarah

Leseproben

1

Building a Unified Data Foundation for Sustainable AI Modernization

These are the usual suspects, but the real answer almost always traces back to infrastructure gaps that exist because of choices made years earlier, often long before anyone was thinking seriously about AI at scale. Legacy systems were built to store and retrieve data for humans, not to feed intelligent agents operating at speed. Data was siloed by department and not designed for cross-system reasoning. Governance frameworks were built around compliance checklists, not the dynamic, high-velocity access patterns that AI demands. These gaps are rarely visible until a promising initiative reaches production, and then they become impossible to ignore. Understanding them is the first step toward building a foundation that is truly ready for what AI requires.

In this chapter, you will:

• Identify the most common challenges organizations face when building AI initiatives
• Understand the architecture of retrieval-augmented generation (RAG) systems
• Explore the core capabilities of AI agents
• Recognize the shared data requirements of these systems, including diverse data types, real-time processing, and intelligent retrieval

Why many AI initiatives stall

Before we can define what a strong data foundation looks like, it's worth understanding why so many AI initiatives fail to get off the ground in the first place.

Take a regional insurance carrier that wants to modernize claims processing by deploying an AI agent. The system is designed to accept a claim, gather all relevant information such as policies, damaged photos, and medical records, and generate a settlement recommendation in minutes instead of days. The organization has access to modern AI models, experienced architects, and the right search and vector capabilities. In testing, the system performed flawlessly but in production, something very different happened.

The initiative unraveled across three fronts. First, policy records, claim histories, and damaged photos were spread across separate systems, forcing the agent to piece results together manually, leading to worse recommendations and eroded trust. Second, the retrieval system surfaced unstructured notes containing sensitive health information that had never been access-controlled, triggering a months-long governance overhaul. Third, vector search infrastructure that handled testing buckled under real workload, stretching settlement times from just a few minutes to fifteen or more.

Similar stories repeat across industries. This is a common pattern seen across teams trying to launch AI initiatives. But the problem isn't with the talent, the models, or the cloud infrastructure. It's the data infrastructure underneath. Teams are often relying on architectural decisions and systems that were built years earlier to solve problems that had nothing to do with AI at scale. Understanding how these gaps create challenges is the first step toward building a data foundation that is ready for what comes next.

The current AI landscape

Understanding the problem is only half the battle. To design a data foundation that holds up in production, we need to know what we're building it for. At present, two AI architectures dominate enterprise adoption: RAG systems and AI agents. They represent the most widely deployed patterns for putting AI to work on real business problems, and they place very specific requirements on the data infrastructure beneath them. Before we can define what a unified data foundation should look like, we need to understand how these systems work and what they're designed to overcome.

How RAG systems retrieve and ground responses

When you interact with a chatbot, such as Claude or ChatGPT, you are interacting with a large language model (LLM), a system trained on vast amounts of public data to understand and generate human-like language. LLMs are powerful, but they have a significant limitation: they don't have access to your company's proprietary data, and they don't automatically learn about new information. Ask one about a specific policy and it may respond with a confident-sounding guess that is simply wrong, a phenomenon known as hallucination.

Figure 1.1 – How a chatbot like Claude generates a response

RAG solves this by first searching your own data for relevant context, then passing that context to the model alongside the original question so the response is grounded in accurate, current information. The quality of that retrieval step is everything.

Figure 1.2 – How RAG systems retrieve relevant data as context to an LLM

RAG systems depend on your data to produce accurate answers, but that only works when the data is prepared and ingested in a way that makes it easy to retrieve the most relevant information. To understand the data and infrastructure this requires, let's consider how this process works.

The following steps walk through how a RAG system works end to end:

1. Identify your data sources: Company knowledge can live in many places, such as databases, documents, emails, chat histories, customer interactions, and more. The first step is knowing where your relevant data lives.
2. Prepare and ingest your data: Once gathered and normalized, your data needs to be converted into vector embeddings. Embeddings capture semantic meaning, so phrases such as "blocked credit card" and "fraud hold" are recognized as related even though they use different words. This requires choosing an embedding model, generating and storing embeddings, and keeping them synchronized as your data changes.

   Figure 1.3 – A diagram illustrating the ingestion stage of a RAG system

Note

Vector embeddings are numerical representations that capture meaning. Modern databases such as MongoDB Atlas store these embeddings as lists of numbers, where each number encodes a specific aspect of the data's meaning.

1. Retrieve the right data: This stage involves querying a knowledge base (such as a vector database) with the user's input to find and return the most relevant documents or passages. To do this really well, you need both semantic and lexical search capabilities, as required by your use case. We'll cover these approaches in depth in , .

   Figure 1.4 – A diagram illustrating the ingestion and retrieval stages of a RAG system

2. Prompt the LLM: The retrieved documents are passed to the model alongside the original question and any additional instructions. The model reads that context and generates its response. The quality of the final answer depends entirely on whether the right documents were retrieved and whether they contain accurate, current information.

   Figure 1.5 – A diagram illustrating the ingestion, retrieval, and generation stages of a RAG system

Each stage of this process, from data preparation to retrieval to response generation, has its own requirements, and those stages have to work together smoothly for the system to deliver accurate, trustworthy answers. That coordination is easy to achieve when your data infrastructure is clean and consolidated but it becomes increasingly difficult as system sprawl grows. The more your data is scattered across disconnected systems, the more places there are for something to break down.

AI agents and autonomous workflows

RAG systems are powerful for retrieval and question answering. But there is a large class of business problems that require systems that can perceive their environment, reason through complex tasks, and take action autonomously. These systems are known as AI agents.

An AI agent is an autonomous system that can accomplish multi-step tasks without manual intervention at each step. By combining an LLM with planning capabilities, tools, and memory, an AI agent can break down complex goals into subtasks, execute actions, and adapt based on the results.

How AI agents work

An AI agent is a system with several interconnected capabilities. First, it perceives. An agent receives input, usually a user request or a triggering event, such as a customer submitting an insurance claim.

Figure 1.6 – An AI agent perceiving an assigned task

Second, it reasons. The agent uses an LLM to understand what the user is asking for and what actions are needed to fulfill that request. It breaks the goal into steps. "To process this claim, I need to: retrieve the policy, assess the damage description, check for any red flags, calculate a preliminary settlement, and draft a recommendation." The agent does not just answer questions about what it should do, it figures out the sequence of actions.

Figure 1.7 – An AI agent reasoning to understand goals and decide the next step

Third, it acts. Unlike a passive chatbot, an agent has access to tools that let it interact with external systems. For example, in insurance claims processing, an agent might retrieve the policy from the claims system, pull damage...