Understanding Retrieval-Augmented Generation (RAG)

About this course

This course focuses on the application of artificial intelligence technologies.

Imagine asking a general AI assistant about your company’s travel policy; you receive a confident, articulate response. The challenge arises when this model has never encountered your specific policy. Retrieval-augmented generation, commonly referred to as RAG, addresses this issue by sourcing the relevant excerpts from your documents and generating an informed answer with citations. This method underpins the functionality of many contemporary document-based AI assistants. However, it is crucial to remember that a citation alone does not confirm accuracy, and the performance of a RAG assistant heavily relies on the quality of the information it accesses.

This course guides you through the inner workings of RAG and helps you evaluate the reliability of its outputs. You will engage with a fictional engineering consultancy as its knowledge manager develops an internal policy assistant, progressing from the initial inquiry to the decision-making phase of a pilot project. The course features diagrams, practical examples, and interactive decision-making activities, without the need for coding, database configurations, or software installations.

Your learning objectives include:

• Differentiating RAG from standalone AI responses, keyword searches, and fine-tuning based on the source of information
• Preparing documents for retrieval: selecting relevant content, cleaning extracted text, chunking appropriately, and tagging each segment
• Describing embeddings and similarity searches in simple terms, and understanding when to merge keyword and semantic matching
• Tracing a query throughout the complete answer development process: retrieval, filtering, ranking, context assembly, generation, and citation
• Identifying the root causes of poor responses: whether due to missing information, an unsuccessful retrieval process, or incorrect evidence usage, and implementing corrective measures
• Verifying that each claim made in a response is adequately backed by the cited passages, including applicable conditions, exceptions, and dates
• Establishing access controls throughout the process and preventing directives buried in documents from functioning as actionable commands
• Creating an evaluation set to independently assess retrieval performance, answer accuracy, evidential support, abstention, and permission compliance

The course culminates with a practical scenario where you will design a new policy assistant and draft a structured pilot brief. You will also compare your brief against an example response and a comprehensive RAG design checklist. Evaluation terminology will adhere to established RAG evaluation frameworks that distinguish retrieval performance from response quality, all while maintaining a vendor-neutral perspective.