Back

Most people don't even know these basics of RAG. I've explained it in a simple way below. 1. 𝗜𝗻𝗱𝗲𝘅𝗶𝗻𝗴 Convert documents into a format that AI can quickly search later. Step-by-step: • 𝗗𝗼𝗰𝘂𝗺𝗲𝗻𝘁: You start with files like PDFs, Word docs, notes, websites, etc. • 𝗘𝘅𝘁𝗿𝗮𝗰𝘁 → 𝗧𝗲𝘅𝘁: The system pulls raw text out of those documents. • 𝗖𝗵𝘂𝗻𝗸𝘀: The long text is broken into 𝘀𝗺𝗮𝗹𝗹 𝗽𝗶𝗲𝗰𝗲𝘀 (chunks). This is important because AI can’t understand very large text at once. • 𝗩𝗲𝗰𝘁𝗼𝗿𝗶𝘇𝗲 / 𝗘𝗻𝗰𝗼𝗱𝗲: Each chunk is converted into numbers called 𝘃𝗲𝗰𝘁𝗼𝗿𝘀. These numbers represent the 𝘮𝘦𝘢𝘯𝘪𝘯𝘨 of the text. • 𝗘𝗺𝗯𝗲𝗱𝗱𝗶𝗻𝗴 𝗠𝗼𝗱𝗲𝗹: A special model does this text → vector conversion. • 𝗦𝗮𝘃𝗲 𝗶𝗻 𝗩𝗲𝗰𝘁𝗼𝗿 𝗗𝗮𝘁𝗮𝗯𝗮𝘀𝗲: All vectors are stored in a 𝘃𝗲𝗰𝘁𝗼𝗿 𝗱𝗮𝘁𝗮𝗯𝗮𝘀𝗲 so they can be searched later. ________________ 2. 𝗥𝗲𝘁𝗿𝗶𝗲𝘃𝗮𝗹 (𝗥) Fetch the most relevant chunks for a user’s question. Step-by-step: • 𝗨𝘀𝗲𝗿 𝘀𝘂𝗯𝗺𝗶𝘁𝘀 𝗮 𝗾𝘂𝗲𝘀𝘁𝗶𝗼𝗻: Example: “𝘞𝘩𝘢𝘵 𝘥𝘰𝘦𝘴 𝘵𝘩𝘦 𝘤𝘰𝘯𝘵𝘳𝘢𝘤𝘵 𝘴𝘢𝘺 𝘢𝘣𝘰𝘶𝘵 𝘵𝘦𝘳𝘮𝘪𝘯𝘢𝘵𝘪𝘰𝘯?” • 𝗤𝘂𝗲𝘀𝘁𝗶𝗼𝗻 → 𝗩𝗲𝗰𝘁𝗼𝗿𝗶𝘇𝗲𝗱: The question is also converted into a vector using the same embedding engine. • 𝗩𝗲𝗰𝘁𝗼𝗿 𝗗𝗮𝘁𝗮𝗯𝗮𝘀𝗲 𝗦𝗲𝗮𝗿𝗰𝗵: The system compares: 1. Question vector 2. Stored document vectors • 𝗠𝗮𝘁𝗰𝗵𝗶𝗻𝗴 / 𝗦𝗶𝗺𝗶𝗹𝗮𝗿𝗶𝘁𝘆 𝗦𝗲𝗮𝗿𝗰𝗵: The database finds chunks whose meaning is closest to the question. • 𝗔𝗽𝗽𝗿𝗼𝗽𝗿𝗶𝗮𝘁𝗲 𝗖𝗵𝘂𝗻𝗸𝘀 𝗢𝘂𝘁𝗽𝘂𝘁: Only the 𝗺𝗼𝘀𝘁 𝗿𝗲𝗹𝗲𝘃𝗮𝗻𝘁 𝗽𝗶𝗲𝗰𝗲𝘀 𝗼𝗳 𝘁𝗲𝘅𝘁 are returned. ________________ 3. 𝗔𝘂𝗴𝗺𝗲𝗻𝘁𝗮𝘁𝗶𝗼𝗻 (𝗔) Enhance the user’s question with relevant information. Step-by-step: • 𝗥𝗲𝗹𝗲𝘃𝗮𝗻𝘁 𝗖𝗵𝘂𝗻𝗸𝘀: The retrieved text pieces are collected. • 𝗠𝗲𝗿𝗴𝗲 𝘄𝗶𝘁𝗵 𝗦𝗼𝘂𝗿𝗰𝗲 𝗖𝗼𝗻𝘁𝗲𝗻𝘁: These chunks are combined into a clean context block. • 𝗣𝗿𝗼𝗺𝗽𝘁 𝗖𝗿𝗲𝗮𝘁𝗶𝗼𝗻: The system builds a new prompt: 1. User’s original question 2. Retrieved context • 𝗔𝘂𝗴𝗺𝗲𝗻𝘁 𝘁𝗵𝗲 𝗣𝗿𝗼𝗺𝗽𝘁: This enriched prompt gives the AI 𝘣𝘢𝘤𝘬𝘨𝘳𝘰𝘶𝘯𝘥 𝘬𝘯𝘰𝘸𝘭𝘦𝘥𝘨𝘦. ________________ 4. 𝗚𝗲𝗻𝗲𝗿𝗮𝘁𝗶𝗼𝗻 (𝗚) Generate a correct, grounded response. Step-by-step: • 𝗘𝗻𝗿𝗶𝗰𝗵𝗲𝗱 𝗣𝗿𝗼𝗺𝗽𝘁 𝗦𝗲𝗻𝘁: The prompt (question + context) is sent to the LLM. • 𝗟𝗟𝗠 𝗠𝗼𝗱𝗲𝗹𝘀 (𝗢𝗽𝗲𝗻𝗔𝗜 / 𝗼𝘁𝗵𝗲𝗿𝘀): The language model reads: 1. The question 2. The retrieved knowledge • 𝗙𝗶𝗻𝗮𝗹 𝗢𝘂𝘁𝗽𝘂𝘁: The model generates a response 𝗯𝗮𝘀𝗲𝗱 𝗼𝗻 𝗽𝗿𝗼𝘃𝗶𝗱𝗲𝗱 𝗰𝗼𝗻𝘁𝗲𝘅𝘁, not guesses. Why RAG Is Powerful? <> Normal LLMs rely only on training data BUT, <> RAG lets LLMs use 𝘆𝗼𝘂𝗿 𝗽𝗿𝗶𝘃𝗮𝘁𝗲 𝗼𝗿 𝗳𝗿𝗲𝘀𝗵 𝗱𝗮𝘁𝗮 and it's easy to update knowledge anytime. ✅ Repost for others so they can understand the very basics of RAG.