Imagine hiring a librarian with

Imagine hiring a librarian with one strict rule: never answer from memory. Only from the books on the shelf.

That is exactly how a RAG chatbot works, and it is my favourite thing to build right now.

Here is the fun part of how it goes:

🔎 You ask a question. “What is the refund policy?”

📖 The librarian runs to the shelf. The system searches your documents by meaning, not just keywords, and grabs the few passages that matter most.

✍️ Then the AI writes the answer, using only those pages. Nothing else.

🤷 And if the shelf has no answer? The librarian says “I could not find that.” I think this is the most underrated feature of all. A bot that admits it does not know is worth more than one that sounds sure and is wrong.

I have built retrieval pipelines like this with OpenAI and Anthropic models, plus a car-dealership chatbot where four AI agents split the conversation between them, a bit like a small team where each person has one job.

A tiny experiment you can try: take any chatbot you use and ask it something it could not possibly know, like a specific detail about a private document. Does it admit it, or does it invent something? You learn a lot about a bot in that one question.

👇 Your turn: what is the funniest or most confident wrong answer a chatbot has ever given you? I want stories.