Generative AI Engineer Interview Questions UK: 8 Real Questions and What Interviewers Want

Most Generative AI Engineer interviews in the UK follow a pattern, and knowing it beats guessing. Below are eight questions from our bank of 15, each with a note on what the interviewer is really after.

Eight Generative AI Engineer interview questions, and what the interviewer is thinking

1. Walk me through how you would design a retrieval-augmented generation system for internal company documents.

Why they ask: RAG is the most commonly built GenAI pattern and this tests end-to-end design.

2. When would you use prompt engineering, RAG or fine-tuning, and how do you decide?

Why they ask: They want judgement about the cheapest effective technique for each problem.

3. How do you evaluate an LLM application, and what are the limits of using an LLM as a judge?

Why they ask: Evaluation maturity separates production engineers from demo builders.

4. What are prompt injection attacks, and how would you defend an LLM application against them?

Why they ask: Security awareness is essential for GenAI engineers, especially with tools and data access.

5. How do you manage cost and latency when running LLMs in production?

Why they ask: Production viability depends on unit economics and responsiveness.

6. Your chatbot gave a customer incorrect information about a refund policy. What do you do?

Why they ask: They want to see incident response, root cause analysis and prevention in an LLM context.

7. Tell me about a time you took a GenAI prototype into production.

Why they ask: They want evidence you can move beyond demos to reliable, supported systems.

8. What do you think makes a strong generative AI engineer, as opposed to someone who just calls an API?

Why they ask: They want to hear your view of engineering rigour in this role.

A sample answer to question 1

Walk me through how you would design a retrieval-augmented generation system for internal company documents.

I would begin with the use case, document types and access rules. Ingestion comes first: extract text, using OCR where needed, clean it, keep metadata such as source, date and permissions, and split it into chunks, typically a few hundred tokens with some overlap, ideally respecting headings. Each chunk is embedded and stored in a vector database such as pgvector, Azure AI Search or Pinecone, often alongside keyword search for hybrid retrieval. At query time, I would rewrite the question if needed, retrieve the top candidates, filter by the user's permissions, rerank with a cross-encoder, and pass the best passages to the model with instructions to answer only from the context and cite sources. I would add guardrails for prompt injection and personal data. The evaluation is critical: a test set measuring retrieval recall, answer faithfulness and relevance, with tools like Ragas or custom judges. Finally, I would log queries and feedback, monitor latency and cost, and set up re-indexing when documents change.

Common mistake to avoid: Skipping evaluation and permissions; treating chunking as an afterthought

Want the rest?

Full model answers, key points and common mistakes for all 15 questions are on the Generative AI Engineer interview questions page. The first five are free, no sign-up needed.