Language models often provide incorrect information because they lack knowledge of specific policies. Retrieval-Augmented Generation (RAG) addresses this by providing the model with relevant documents before it answers. This guide explains RAG using a library analogy: a good librarian retrieves specific pages rather than relying solely on memory. RAG involves retrieving relevant passages, augmenting the prompt with them, and generating an answer based on that information.RAG is beneficial when dealing with large or frequently updated documents, or when cost efficiency is important. Instead of pasting entire documents into every prompt, RAG allows for targeted document retrieval. The article outlines building a simple RAG system for a shop's help center using Python and SQLite. This system will chunk documents, index them for full-text search, retrieve top passages, construct a prompt, and verify the answer's citations.The implementation involves a chunk_markdown function to divide documents into meaningful passages, preserving headings for context. The build_index function creates a SQLite FTS5 virtual table to store these passages and enables efficient searching. The search function processes a question, expands it with synonyms, removes common words, and queries the index using a robust method.The build_prompt function formats the retrieved passages as numbered sources for the language model. A check_answer function performs basic validation to ensure the model cites existing sources and does not invent numbers. The core answer function orchestrates these steps, first searching for relevant passages and only calling the language model if relevant information is found. A placeholder extractive_model is provided for testing without a real language model API.