Why it matters
AI grounding links each claim to retrievable source passages. It supports trust, debugging, and compliance review. Ungrounded confident sentences are liabilities in customer chat.
How it works
Map generated sentences to chunk IDs, validate mappings in post-processing, surface citations in UI, fail tests when unsupported sentences appear.
Example
Nintendo grounds refund window statements to Refund Policy section 2.1 passages shown inline. Unsupported comfort language gets stripped before the answer ships.
Common mistakes
- 1Citations that do not support the sentence they decorate
- 2Grounding only in UI footnotes the model ignored during generation
- 3No automated check that every numeric claim has a source
Definitions are useful. Governed knowledge is better.
Wiki helps you turn these concepts into a real system for the knowledge behind your AI.
See Wiki in action