This scenario requires understanding visual content from a photo and then generating a textual explanation. That means the system consumes one modality as an image and produces another modality as text. This cross-modality capability is exactly what a multimodal approach provides, since it jointly handles vision and language to produce coherent natural language output based on visual input.

The primary benefit of RAG (Retrieval-Augmented Generation) in this context is its ability to ensure the chatbot provides accurate and up-to-date information. By retrieving relevant and recent policy documents from Cloud Storage in real-time and then grounding the LLM's response with this information, the chatbot avoids hallucinating or providing outdated answers, which is crucial for an internal knowledge base.

The correct option is Human in the Loop (HITL) because a licensed attorney must review, edit, and approve each AI draft before any client sees it.
Adding human oversight ensures accuracy, legal compliance, and accountability for outputs. HITL is a recommended practice for high-stakes domains where expert judgment must validate model results before release, which matches a legal drafting workflow that requires professional approval.

The objective is to make the generated summaries concise--that is, to control their length. In the configuration of a generative AI model, particularly a large language model (LLM), the parameter used to directly control the maximum size of the response is the Output Length parameter (often referred to as max_output_tokens or max_tokens). By setting a low limit on this parameter, the team can ensure that the model is forced to terminate its response once that limit is reached, resulting in a shorter, more concise summary that is "easy to digest," as requested.

Adversarial robustness needs to be designed into the model from the start and then sustained in production. During training you can harden models with adversarial training, robust data augmentation, regularization, and careful evaluation against adversarial and out of distribution test sets. In production you should continuously monitor for drift, anomalies, and suspicious input patterns and you should feed incidents back into retraining so the system improves over time.
This lifecycle approach ensures protections evolve with attacker tactics and with data and model changes.