8. RAG, Hallucination, and Verification

The lesson

Hallucination means the AI produces plausible text that is not supported by the facts. RAG reduces this by giving the model documents to quote or use.

For business, legal, electrical, medical, or production code decisions, always ask for sources, inspect the source, and test the result.

TPEE's Data and Examples sections are where you tell the AI which facts matter most.

A procurement officer builds a prompt that answers warranty questions only from the supplied contract extracts, asking for the exact clause and a clear flag when the text does not cover the question.

A community pharmacist grounds an answer in the supplied dispensing guidelines, requiring a source line for every claim and an explicit note where the documents are silent.

A site engineer supplies the defect-liability extracts as context and asks the model to quote the clause as written, marking anything the documents do not support.

A reference librarian answers reader questions only from the catalogue and policy excerpts on hand, and says plainly when the material does not cover the query.

A laboratory scientist asks for a method summary based on the supplied procedure and requires a section reference for each step, rejecting additions that cannot be traced.

A newsroom researcher provides the public records as data and asks the model to mark any sentence that cannot be traced back to those records.

A bank compliance analyst supplies the internal policy extracts and requires the model to quote the paragraph behind each conclusion before the note is circulated.

A non-profit grants officer grounds a funding-summary prompt in the supplied programme rules, asking the model to identify which rule supports each statement it makes.

A property manager feeds the tenancy agreement extracts into the prompt and asks for a sourced answer, flagging any point the agreement does not address.

A security analyst provides the incident report and the log excerpts, then asks the model to separate confirmed findings from assumptions and unknowns.

Check yourself