The correct option is Hybrid conversational agent.

This design combines deterministic dialog management for routine FAQs with large language model based responses for open ended or nuanced questions. You can implement strict guided dialog and fixed replies for predictable queries to ensure consistency, then defer to an LLM to interpret intent and generate natural responses when the question falls outside predefined paths. This satisfies both the need for guaranteed answers on known topics and the need for flexibility on unfamiliar requests.

Purely Generative Agent is not appropriate because it relies entirely on an LLM which cannot guarantee fixed phrasing or strictly controlled answers for routine FAQs. That would undermine the requirement for consistent guided dialog.

Retrieval-augmented generation pipeline focuses on grounding model outputs with external data, which can help accuracy, yet it does not by itself provide the strict guided dialog with fixed replies that the team requires. It is a technique rather than a complete agent pattern that blends deterministic flows with generative responses.

Purely Deterministic Agent cannot handle unfamiliar or nuanced questions with natural language flexibility since it depends only on predefined intents and responses. This would fail to meet the requirement for more natural responses when the user asks novel questions.

Full AWS Practitioner Certification Question