Reducing hallucinations
Ground the model in verified sources with RAG, require it to cite the passage it used, and evaluate retrieval quality. Together these cut hallucinations 70–90%, far more effective than prompt-engineering a model to be careful.
Grounding and citations
The reliable fix is architectural: give the model the right source at query time and make it cite what it used, so answers are checkable. Prompting a model to "be accurate" does not add facts it never had.
Requiring citations does two things: it forces the answer to trace to a real source, and it makes wrong answers visible because the cited passage will not support them. Combined with grounding and retrieval evaluation, this turns an unverifiable assistant into one a subject-matter expert can spot-check. The residual hallucinations that remain almost always trace back to retrieval surfacing the wrong context, so fixing retrieval is fixing hallucination.
Frequently asked questions.
Can better prompts stop hallucinations?
Only marginally. Prompts cannot give the model facts it never had. The durable fix is grounding it in your verified data with RAG and requiring citations so answers are checkable.