Scenario

You are asked to design a GenAI assistant for a cloud platform. The first version should answer questions about product docs, deployment guides, pricing, and operational runbooks, and later support guided actions such as troubleshooting and setup recommendations. Users include developers, platform engineers, and support teams. The team wants a design that can run on Crusoe Cloud and can grow from a retrieval-based assistant into a more capable workflow over time.

Constraints

• Responses should feel interactive for common questions
• The system must keep hallucinations low on factual product answers
• Retrieved content may contain stale, conflicting, or malicious instructions
• The design should be cost-aware enough to support production traffic

Available Resources

• Crusoe Cloud compute for model serving, indexing, and batch jobs
• Product documentation, runbooks, support articles, and API references
• Access to a hosted LLM API and open-source embedding models
• A small team that can label a limited evaluation set before launch

Question

How would you design a GenAI solution on Crusoe Cloud, including how you would ground answers, evaluate quality before launch, and handle safety risks such as hallucination and prompt injection?