2026 年 AI 工程師面試問題：值得準備的 5 個答案

AI engineer interviews in 2026 mostly ask you to reason about a working LLM system: why a RAG answer is wrong, how a tool call executes, what a chatbot costs per month, how to keep a model from being tricked. Here is the condensed version of a 13-question guide and the answer shape interviewers reward.

What companies actually publish

Very little. Google says its process typically takes roughly 6-8 weeks, assesses candidates against clear rubrics, and does not permit AI tools in interviews. Anthropic says live interviews are "all you", with no AI assistance unless it says otherwise. Most round-by-round breakdowns online are candidate reports.

The rounds fall into four buckets: concepts (tokens, embeddings, tool calling, fine-tuning trade-offs), practical coding (call an LLM API, validate output, handle failures), system design (RAG or agents with cost, evals and failure modes) and behavioral (an AI project you shipped and what you'd change).

Five answers worth having ready

1. Tokens. OpenAI's rule of thumb for English is about 4 characters, or 0.75 words, per token. Tokens drive cost (input and output are priced separately), limits (the context window is a token budget) and odd behaviour (models see tokens, not letters, so counting letters fails, and non-English text often costs more).

2. Why you can't unit-test an LLM feature. Correct answers vary in wording, and the same prompt can return different text, so exact-match assertions are flaky and fail good answers. The replacement is an eval set: fixed realistic inputs, a scoring method and a pass rate you re-run whenever the prompt, model or retrieval changes.

3. RAG gives confident wrong answers. Split the pipeline before touching the prompt. For each failing question, check whether the right passage was retrieved at all (if not: chunking, embeddings, query phrasing). If it was, give the model only that passage. If it still fails, the fault is generation.

4. Estimating chatbot cost. State the formula first: conversations × turns × (input tokens × input price + output tokens × output price). Then name what people forget: the system prompt, retrieved passages, tool definitions and the full history are re-sent every turn, so cost grows faster than linearly with conversation length.

5. Handling 429s and timeouts. Honor Retry-After, fall back to exponential backoff with jitter, and cap attempts, single waits and the total deadline. Retry timeouts, 429s and 5xx errors, never bad requests or auth errors. Anthropic documents a 429 for a monthly spend cap that keeps failing until access resumes, so not every 429 is retryable.

Prompting, retrieval or fine-tuning?

Match the technique to the failure. Misunderstood task or format: better prompt and structured output. Missing or changing facts: retrieval. Consistent style or a narrow skill the prompt can't teach: fine-tuning. Go in that order, because each step costs more to build and maintain, and have the eval set first so you can prove the cheaper option wasn't enough.

Prepare by level

• Junior: correct mental models (tokens, the tool loop, why evals exist, prompt injection). Build one small RAG app with a 20-question eval set.
• Mid: diagnose failures by stage and argue trade-offs with numbers, including a worked cost estimate for something you shipped.
• Senior: limit blast radius, treat evals as release gates, and know when not to use a model.

The full guide has all 13 questions with follow-ups: AI Engineer Interview Questions (2026) on StackChief.

This post was written with the help of AI, as a summary of the StackChief article above.