Interview Questions

In an interview, they want to see that you think about the whole system, not just the model. Here are the questions that come up most, with a strong way to answer each one.

• Think about the system
• Common questions
• Practise out loud

How would you build a support bot?

A strong answer starts with the goal, and what mistakes cost. Then describe the pipeline. Redact private data, triage with a cheap classifier, retrieve policies using hybrid search and reranking, and generate a grounded reply with citations. Add tools for live data, guardrails on the output, and a human hand off. Finish with evaluation using a golden set, and launch in shadow mode.

Your RAG answers are poor. What do you do?

Show a method. First, separate retrieval from generation. Take failed questions, and see if the right text was found. If not, fix chunking, add keyword search, or a reranker. If it was found, but the answer's wrong, fix the prompt, or tell the model to use only the sources. Measure each change using a golden set. Interviewers love that order.

Fine-tune or use RAG?

Be direct. RAG adds facts, stays up to date, and it can cite sources. Fine tuning changes style, format, and behaviour. I'd try prompting first, then RAG, and fine tune only if I needed a consistent behaviour that prompts couldn't deliver, or I wanted to cut costs with a smaller model. Many systems use both.

How do you control LLM costs?

Give a list. Count tokens. Route easy cases to small models. Cache repeated questions. Keep prompts and retrieved text short. Use batch APIs for offline work. And set budgets and alerts. Mention that you'd estimate costs before launch, which shows real experience.

Summary