Source attribution

In Prompt Engineering you designed the citation schema — source IDs plus verbatim quotes, mechanically checkable. In a RAG pipeline that design stops being a prompt pattern and becomes architecture: every answer flows through a verification gate before anyone sees it.

def verify_citations(output: dict, chunks: dict[str, str]) -> dict:
    """Returns the output annotated with per-citation verification."""
    for c in output["citations"]:
        source_text = chunks.get(c["source_id"], "")
        c["verified"] = _normalize(c["quote"]) in _normalize(source_text)
    output["all_verified"] = all(c["verified"] for c in output["citations"])
    return output

def _normalize(s: str) -> str:   # whitespace + quote-mark tolerance
    return " ".join(s.split()).replace("\u201c", '"').replace("\u201d", '"').lower()

• Normalize before matching. Models faithfully copy text but harmlessly vary whitespace and curly quotes; a naive exact match fails honest citations. Normalize both sides; keep the check strict on words.
• A failed verification is a routing signal, not a formatting bug. It means the model asserted something its sources don't contain — exactly the event your fallback ladder exists for: retry with the violation named, or degrade to 'show sources without the claim'.
• Attribute per claim, not per answer — you know this rule; the pipeline now enforces it because unattributed sentences are detectable (a sentence with no adjacent [S#] in a fact-bearing answer).

Citation UX: where trust is actually won

• Cite with display metadata, not internal IDs: 'Holiday Returns → Extended window', linked to the source. chunk_147 impresses nobody.
• Show the quote on hover/expand. Users who can check but don't have to are the users who trust you. The checkable-but-unchecked citation is the entire psychology of RAG credibility.
• Don't over-cite. Four citations on one plain sentence reads as noise. One precise citation per claim, deduplicated per source.
• When `answerable` is false, show the nearest sources anyway — 'I couldn't find this in the docs; closest material:' turns a dead end into a hand-off. Users forgive 'not found'; they don't forgive fabrication.