Lesson 5 of 8

Ground answers in sources with citations and 'I don't know'

Complete the RAG pipeline: generate answers that cite their sources, refuse to hallucinate when retrieval is thin, and surface provenance so learners trust the output.

You're on lesson 5 of 6 in the free RAG module. Unlock the full AI Engineer curriculum →

Step 1 · concept

Ungrounded answers are a black box

A grounded answer is one where every claim traces back to a source document. A grounded RAG system doesn't just answer the question — it shows its work.

This matters for three reasons:

  • Trust. Users can verify claims against the original source. "The model said X" becomes "Document Y says X", which is a very different guarantee.
  • Debugging. When the answer is wrong, citations tell you whether the problem was retrieval (wrong source returned) or generation (right source, misread). Without grounding, you stare at a black box and guess.
  • Accountability. In regulated domains — healthcare, legal, finance — you need an audit trail showing where information came from. An ungrounded answer is unshippable in those contexts.

Without grounding, your RAG system is a black box: the user gets an answer but has no way to assess reliability. This lesson closes that gap with three production patterns: source-labelled context, the strict-refusal branch, and structured citation output.

Your RAG chatbot answers a customer's refund question incorrectly. The generated answer claims 'refunds are processed within 48 hours' — the real policy is 14 days. Why does grounding matter MORE than just fixing the answer?