Perfect retrieval can still produce an embellished answer — the model receives the right evidence and adds a detail of its own anyway. The two answer-side checks on a chat's Context tab close that gap: one requires answers to cite the supplied evidence, the other verifies the claims themselves. Both are off by default and configured per chat.
- Citation Integrity — every factual claim in the answer must carry a citation that resolves to one of the retrieved chunks. Answers whose citations don't check out are replaced with "I could not verify a supported answer from the supplied sources." rather than shown. Valid citations show to visitors as numbers such as [1], in the order of the source cards.
- Strict Answer Verification — after the answer is generated, the accuracy model checks each claim against the evidence. Unsupported claims trigger one rewrite; if the rewrite still doesn't verify, the visitor gets the same "could not verify" message instead.
- Expected accuracy calls — the per-message cost of the current switches. Verification adds one support check and at most one regeneration.
Which providers support citations #
Citation Integrity uses provider-native citation features and currently works when the chat's answering LLM profile is OpenAI or Anthropic. With any other provider the switch stays quietly inactive — the chat answers normally, without enforcement, and Diagnostics records the mismatch so you can spot it.
Anthropic profiles get an extra benefit: citations arrive as precise pointers into the exact evidence passages, and repeated evidence is cached across requests where eligible.
What a citation does — and doesn't — prove #
A validated citation proves the answer is pointing at real, supplied evidence — not an invented source. It does not by itself prove every sentence is true; that's what Strict Answer Verification adds. Use Citation Integrity alone for a good balance, and add verification for chats where a wrong answer is worse than no answer (policies, pricing, compliance).
Two practical notes:
- Both checks buffer the answer before display, so streaming chats show the reply all at once when these are on — the visitor never sees an answer that's about to be retracted.
- The replacement text is the chat's Fallback Message (the same one used for no-results). Write it so it invites rephrasing: "I couldn't verify an answer from the documentation — try asking a more specific question."
If visitors report getting the fallback message often, see the troubleshooting checklist — it's usually a threshold or evidence-coverage issue, not a bug.
What to do next #
Measure the effect: the Retrieval Lab shows whether your grounded chats are answering from the right sources before and after these switches.
