Prompt engineering to curb extraction hallucinations

WHAT THIS TESTS Whether you can apply concrete prompting techniques to extraction and, just as importantly, state where prompting stops being enough.

A GOOD ANSWER COVERS PROMPTING TECHNIQUES. Tell the model to extract only information explicitly present in the source and to output null or not found rather than inventing a value for missing fields, which removes the pressure that drives confabulation. Constrain the output to a strict schema, ideally enforced with JSON or structured-output decoding, so the model fills fixed slots instead of free-forming. Provide few-shot examples that resemble the real documents, including examples where some fields are genuinely absent so the model learns to emit null. Ask the model to attach the verbatim source span for each extracted value, which both improves grounding and gives you a check. Lowering temperature reduces some random invention. LIMITATIONS. Prompting changes the distribution but offers no guarantee; the model can still confidently fabricate, especially on document layouts unlike your examples. Prompts are brittle and can regress when documents shift. They do not verify correctness. For dependable extraction you need validators that reject malformed or unsupported outputs, retrieval grounding, and ideally a second-pass verification that confirms each value appears in the source.

COMMON WRONG ANSWERS Claiming a clever prompt eliminates hallucination. Not allowing a null option, forcing the model to guess. Omitting a schema. Ignoring that prompting needs an external verification layer. Assuming test-set success generalizes to all new documents.

LIKELY FOLLOW-UPS Why does allowing null reduce hallucination? How would structured-output decoding help? What verification layer would you add beyond prompting? How do you detect regressions on new document types?

ONE CONCRETE EXAMPLE Extracting invoice_total from a document lacking that field, a naive prompt makes the model guess a plausible number. Revised, the prompt says return null if the value is not explicitly present and attach the source text, plus a few-shot example with a missing total mapped to null. The model now returns null with no span. A downstream validator confirms each non-null value's span actually exists in the source, catching any residual fabrication that prompting alone missed.