IBM's Katie McDonald on AI: ADK vs. RAG

In a concise explanation, IBM Demand Strategist Katie McDonald breaks down the fundamental differences between two key AI agent architectures: Agent Development Kits (ADK) and Retrieval Augmented Generation (RAG). McDonald, whose role at IBM focuses on strategic demand generation for AI solutions, offers a clear framework for understanding when to employ each approach, and when a hybrid model might be most effective.

Understanding the Architectures

McDonald begins by drawing a relatable analogy to hardware stores. One aisle is stocked with tools for performing tasks, and another with reference guides and information. This metaphor serves to illustrate the core functions of ADK and RAG.

ADK (Agent Development Kit): McDonald defines ADK as being focused on action and reasoning. AI agents utilizing an ADK are designed to perform multi-step tasks, follow instructions, and make decisions. This approach is suitable when the AI needs to execute processes, manage workflows, or provide assistance in areas like IT or HR. Examples include automating tasks, drafting content, or supporting with task coordination.

The full discussion can be found on IBM's YouTube channel.

RAG (Retrieval Augmented Generation): In contrast, RAG is centered on knowledge and accuracy. RAG-based AI systems are built to retrieve information from external data sources and present it accurately. This is ideal for situations where the AI needs to answer questions based on specific documents, such as PDFs, policies, technical documentation, or knowledge bases. RAG is particularly useful when the AI needs to access and synthesize information that is not inherently part of its training data.

When to Use Which Approach

McDonald clarifies the decision-making process by posing a simple question: Does the AI need to 'do something' or 'know something'?

ADK is the preferred architecture when the AI's primary function is to perform actions, follow a sequence of steps, and make decisions. This is akin to an AI that needs to execute a process or manage a workflow. The value here comes from the AI's reasoning capabilities and its ability to act autonomously based on predefined rules or learned behaviors.

RAG, on the other hand, is best suited for applications where the AI needs to access and leverage specific, often proprietary, information. When the AI's output must be grounded in factual data from documents or knowledge bases, RAG ensures that the responses are accurate and contextually relevant. This is crucial for tasks like answering questions about specific company policies or summarizing technical documents.

The Power of Hybrid Systems

McDonald highlights that the most sophisticated and effective AI systems often combine both ADK and RAG principles. This hybrid approach allows AI agents to both perform complex actions and access external knowledge to inform those actions.

For instance, an AI assistant that helps draft legal documents might use RAG to retrieve relevant case law and statutes, and then use ADK principles to structure the document, follow specific formatting rules, and perform other procedural steps. Similarly, an IT support AI could use RAG to find solutions in a knowledge base and then use ADK to guide the user through troubleshooting steps.

The video concludes by emphasizing that understanding the core requirements of the AI task—whether it's primarily about action or information retrieval—is key to selecting the appropriate architecture, or to designing a hybrid system that capitalizes on the strengths of both ADK and RAG.