Point it at a document a general chatbot has never read, such as a contract, a policy, or a research paper, and get answers grounded in that source: every claim cited, every claim checked against a real passage, and an honest “I don’t know” when the document does not say.
Why not just ask ChatGPT?
A general chatbot answers from memory and will fill any gap with a confident, plausible-sounding guess. That is fine for trivia and dangerous for a contract, a policy, or a research finding. Grounded answers only from the document you give it, shows the exact passage behind every sentence, and refuses when the document does not hold the answer. Load a sample above, or bring your own document, and check the work yourself.
Answers are built only from the document you load, so they cannot quietly drift into training-data guesses.
A second model checks each sentence against a real passage and flags anything it cannot find in the source.
Ask something the document does not cover and it declines, rather than inventing a confident wrong answer.