The correct option is Fine tune the model on a curated corpus of legal documents.

Fine tune the model on a curated corpus of legal documents adapts the foundation model to the writing style, structure, and vocabulary of the legal domain. On Vertex AI this aligns the model weights with your examples so it consistently produces outputs that follow legal conventions and use specialized terminology, which is exactly what the startup requires. Fine tuning reduces the amount of prompt engineering needed for style enforcement and yields more reliable domain adherence across varied prompts.

Retrieval augmented generation with a private legal knowledge base improves factual grounding by injecting relevant documents at generation time, yet it does not change the underlying model behavior or style. It helps the model cite the right information but it will not reliably enforce legal tone and terminology across tasks without fine tuning.

Human in the Loop review for every response can catch mistakes and improve quality assurance, but it does not specialize the model and it scales poorly due to cost and latency. It is better used to curate training data and to evaluate a fine tuned model rather than as the primary method for domain adaptation.

Adjust the model temperature to a lower value only controls randomness and diversity of outputs. Lower temperature can make responses more deterministic, but it does not teach legal vocabulary or conventions. Without fine tuning the model will not reliably produce domain specific writing even at a low temperature.

Full AWS Practitioner Certification Question