A startup develops a generative AI tool for legal document review. To ensure explainability, they want the model to cite specific clauses from source documents when making assertions. Which technique should they use?
Grounding retrieves passages from the source documents and injects them into the model's context, so each assertion can be tied to a specific clause. This directly satisfies the explainability requirement by making citations traceable to retrieved text rather than relying on parametric memory, which cannot attribute claims to sources.
Why this answer
Grounding with a retrieval system (RAG) provides the model with actual source documents at inference time, so it can cite specific clauses verbatim from those sources rather than generating citations from memory. This directly supports explainability because every assertion can be traced back to a retrieved passage. Fine-tuning and prompting alone cannot guarantee accurate citations because the model may hallucinate references.
Exam trap
Generative AI Leader often tests the misconception that fine-tuning or prompt engineering alone can produce reliable citations, when grounding via retrieval is required to cite actual source content.
How to eliminate wrong answers
Option A is wrong because fine-tuning on citation examples teaches the model a citation style but does not give it access to the actual source documents at inference time — it can still hallucinate clause references. Option B is wrong because chain-of-thought prompting improves reasoning transparency but does not provide source documents, so citations would still be generated from parametric memory. Option D is wrong because prompt engineering asking for citations cannot force the model to cite real clauses it has not been given — it will fabricate plausible-looking references.