Part 1 of 6 · 1 min read
Overview
In this module you will
- Define grounding and say what a citation does and does not prove.
- Build a prompt with numbered sources, answer rules and a fixed refusal sentence.
- Write a sentence-level citation checker and say exactly which failures it can and cannot catch.
- Add a gate that refuses to answer when the retrieved passages cannot support one, and measure its false refusals.
- Explain why a perfectly cited answer can still be wrong.
An agent that answers from retrieved text is only useful if people (and other programs) can trust and check what it says. A support agent that quotes the wrong price, or invents a policy and sounds sure, costs real money. Citations let a reader click through to the source. Refusals stop the agent from filling gaps with plausible guesses. And a checker means a wrong citation is caught by code instead of by an angry customer. This module builds all three in a small, runnable form.
About the examples
The documents are an invented 10-passage help centre and the 24 questions were written by me (18 answerable, 6 not). The "model" is a rule program that copies the best-matching source sentence and cites it; it never says "I don't know" on its own, which is how a real model behaves when told to answer. "Failures" such as a wrong number are produced by code that damages honest answers on purpose. A real model fails in more varied, subtler ways. The checker is word-matching, a simple stand-in for the entailment model or judge a real system would use. The numbers show mechanism and direction, not benchmarks. No real model or service was called.