FDE Interview Prep · FI5
Rapid-Fire Review: 42 Questions to Know Cold
Most FDE loops include ten minutes where the interviewer stops discussing designs and starts firing short questions: what's a KV cache, why hybrid search, what does idempotent mean. Nobody expects an essay. They want a crisp, correct two-sentence answer, delivered without hesitating. This page is the review sheet for that round: forty-two questions with short answers, each linked to the lesson that teaches it properly.
⏱️ ~1 hour⚡ 42 questions🎯 Essential→Tech-lead
Learning objectives
- Answer core LLM, retrieval, agent, evaluation, inference, and systems questions in two or three sentences.
- Recognise when a short answer is complete, and stop there instead of rambling.
- Handle a question you don't know honestly, by reasoning out loud instead of bluffing.
- Know which lesson to revisit for every topic you can't yet answer cold.
1 · How rapid-fire rounds work essential
The rapid-fire round catches out strong engineers more often than the design round does. After forty minutes of open-ended discussion, the interviewer switches gear: ten to fifteen short questions in about ten minutes. Each one checks a single fundamental. Candidates who are excellent at systems thinking sometimes stumble here, either because they launch into a five-minute answer to a thirty-second question, or because they've used a technique for years without ever putting it into words.
The round is a fluency check, not a trick. An FDE has to explain these ideas to clients, security teams, and executives with no notice, so the interviewer wants to see that the fundamentals are solid enough to say plainly. The format rewards a specific rhythm: give the definition in a sentence, give the why it matters (or a tradeoff) in a second sentence, and then stop. If they want more, they'll ask. Stopping cleanly reads as confidence; padding reads as uncertainty.
And when you don't know an answer, don't bluff. Interviewers in this round know the material well and spot invented answers immediately, which costs you more than the gap itself. Say "I'm not sure, but here's how I'd reason about it," and then reason from what you do know. That turns a gap into a demonstration of the very skill FDEs use every week: working out an unfamiliar system from first principles.
Define → why → stop. Two sentences answer most rapid-fire questions completely. When you hit a gap, reason out loud from what you know instead of inventing an answer.
🗺️ How to read this diagram
- Read the top row left to right: a one-sentence definition, then one sentence on why it matters or what it trades off.
- The green "Stop" box is the step people skip. A complete short answer followed by silence is a strong signal, not an awkward one.
- The amber bar is the fallback for a question you can't answer: say so, then reason from first principles.
In short: two sentences, then stop. If you don't know, reason out loud instead of guessing.
How to use this pageKeep each answer closed. Read the question, say your answer out loud in two or three sentences, then open it and compare. Any question where you hesitated, rambled, or got it wrong: follow the Full lesson → link and come back to it tomorrow. Aim to get through all forty-two without opening an answer early.
2 · LLM fundamentals essential
⚡ Rapid-fire
What is a token, and what is the context window?
Show answer
A token is the unit a model reads and writes — usually a word piece, roughly three-quarters of an English word on average. The context window is the maximum number of tokens the model can handle in one request, counting the input and the output it generates, and it's also what you pay for. Full lesson →
⚡ Rapid-fire
What do temperature and top-p control?
Show answer
Both control how the next token is sampled from the model's probability distribution. Temperature rescales the distribution — low values make the likeliest tokens dominate, high values flatten it so output is more varied — while top-p samples only from the smallest set of tokens whose probabilities add up to p. Some newer models fix these and expose a reasoning-effort setting instead. Full lesson →
⚡ Rapid-fire
Why do LLMs hallucinate, and what actually reduces it?
Show answer
A model generates the most plausible continuation, not a verified fact, so when it lacks the information it can produce fluent but false text. What helps is grounding answers in retrieved sources and requiring citations, allowing and encouraging "I don't know," checking outputs against the sources, and measuring the hallucination rate with evals. Full lesson →
⚡ Rapid-fire
When do you use prompting, RAG, or fine-tuning?
Show answer
Start with prompting; it's the cheapest and fastest to change. Use RAG when the model needs knowledge it doesn't have, especially private or frequently changing information. Use fine-tuning to change behaviour, format, or style consistently — not to teach facts, which go stale and are hard to update in weights. Full lesson →
⚡ Rapid-fire
What's the difference between the system prompt and the user message?
Show answer
The system prompt sets the assistant's role, rules, and constraints for the whole conversation, and is written by the developer. The user message is the per-turn input. Keeping trusted instructions in the system prompt and untrusted content in user turns is both cleaner and a basic defence against prompt injection. Full lesson →
⚡ Rapid-fire
What is an embedding, and why use cosine similarity?
Show answer
An embedding is a vector that represents the meaning of a piece of text, so that texts with similar meaning end up close together. Cosine similarity compares the angle between two vectors and ignores their length, which makes it a good measure of "same meaning" regardless of text length; for normalised vectors it equals the dot product. Full lesson →
⚡ Rapid-fire
What does it mean to treat the context window as a budget?
Show answer
Every token in the window costs money and latency, and models use information less reliably when it's buried in a long context. So you spend it deliberately: rank and trim retrieved content, summarise old history, and keep stable content at the start so it can be cached. Full lesson →
3 · Retrieval & RAG intermediate
⚡ Rapid-fire
Why does chunking matter in RAG?
Show answer
You retrieve chunks, not documents, so the chunk is the unit of meaning the model sees. Chunks that are too large dilute the match and waste context; chunks that are too small lose the surrounding context they need. Splitting on natural boundaries like headings and sections, with a little overlap, usually beats fixed-size splits. Full lesson →
⚡ Rapid-fire
What is hybrid search, and how do you combine the results?
Show answer
Hybrid search runs keyword search (such as BM25) and dense vector search side by side, because keywords catch exact terms like IDs and names while vectors catch paraphrases. The scores aren't on the same scale, so you combine by rank, typically with Reciprocal Rank Fusion: each document scores the sum of 1/(k + rank) across the lists. Full lesson →
⚡ Rapid-fire
What does a re-ranker do, and what does it cost?
Show answer
A re-ranker — usually a cross-encoder — reads the query and each candidate passage together and scores relevance much more accurately than the first-pass retriever. It's too slow to run over the whole corpus, so you re-rank a shortlist of, say, the top 50, at the cost of extra latency per query. Full lesson →
⚡ Rapid-fire
What is approximate nearest neighbour (ANN) search, and why HNSW?
Show answer
Exact nearest-neighbour search compares the query with every vector, which is too slow at millions of vectors. ANN indexes trade a little recall for a large speed-up; HNSW is a popular one that builds a layered graph you can navigate to close neighbours in roughly logarithmic time, at the cost of extra memory. Full lesson →
⚡ Rapid-fire
How do you tell whether a bad RAG answer is a retrieval problem or a generation problem?
Show answer
Look at what was retrieved. If the correct passage isn't in the retrieved set, it's a retrieval or ingestion problem; if it is there and the answer is still wrong, it's a generation or grounding problem. Measure the two separately — for example recall@k for retrieval and faithfulness for generation. Full lesson →
⚡ Rapid-fire
Why use metadata filters in retrieval?
Show answer
Filters restrict the search to documents that are allowed and relevant — by tenant, permission, date, or version — before ranking. They improve precision and, more importantly, enforce access control, so one customer's or one user's documents can never leak into another's answer. Full lesson →
⚡ Rapid-fire
What is query transformation?
Show answer
Rewriting the user's query before retrieval so it matches the documents better — for example expanding abbreviations, splitting a multi-part question into sub-queries, or generating a hypothetical answer to search with. It helps when users phrase things very differently from how the source documents are written. Full lesson →
4 · Agents & tools advanced
⚡ Rapid-fire
How does a tool-calling loop work?
Show answer
You send the model a list of tools with schemas. When it decides to use one, it returns a tool call with arguments; your code runs the tool and sends the result back, and the model continues. The loop repeats until the model returns a final answer instead of another tool call — the model never runs anything itself. Full lesson →
⚡ Rapid-fire
What is MCP?
Show answer
The Model Context Protocol is an open standard for connecting AI applications to tools and data. A server exposes tools, resources, and prompts once, and any MCP-compatible client can use them, so you don't write a custom integration for every pairing of app and system. Full lesson →
⚡ Rapid-fire
Why cap an agent's steps and spend?
Show answer
A confused agent can loop — retrying the same failing tool or wandering — and every step costs tokens, time, and possibly side effects. Hard limits on steps, tokens, and wall-clock time bound the worst case, and hitting a limit should end the run cleanly or hand it to a human. Full lesson →
⚡ Rapid-fire
When should you use a workflow instead of an agent?
Show answer
When the steps are known in advance. A workflow — fixed steps that call the model where needed — is cheaper, faster, and far easier to test. Reach for an agent only when the path genuinely depends on what's discovered along the way, and the value justifies the extra cost and unpredictability. Full lesson →
⚡ Rapid-fire
What makes a tool easy for a model to use correctly?
Show answer
A clear name, a description that says when to use it (and when not to), and a tight input schema — enums instead of free text, required fields, sensible types. Tools shouldn't overlap in purpose, and they should return concise, structured results with errors the model can act on. Full lesson →
⚡ Rapid-fire
How do you make agent actions safe?
Show answer
Validate every tool argument in your own code before acting, give tools least privilege, and require human confirmation for irreversible or high-impact actions. Log every call for audit. Never rely on the prompt alone for safety, because prompts can be overridden. Full lesson →
⚡ Rapid-fire
Why are agent traces essential?
Show answer
Agents fail in multi-step ways that are invisible in the final answer. A trace records each step — the prompt, the tool calls, their arguments and results, latency and tokens — so you can see exactly where a run went wrong and turn real failures into eval cases. Full lesson →
5 · Evaluation & safety professional
⚡ Rapid-fire
What is a golden set?
Show answer
A fixed, curated set of realistic inputs with known good answers or required facts, used to measure quality the same way every time. It turns "seems fine" into a number you can track across prompt, model, and data changes. Full lesson →
⚡ Rapid-fire
What is LLM-as-judge, and what's the catch?
Show answer
Using a model to grade outputs against a rubric, which scales evaluation to things simple checks can't measure, like groundedness or helpfulness. The catch is that the judge has its own biases — towards longer answers, towards the first option, or towards its own style — so you validate it against human grades before trusting its scores. Full lesson →
⚡ Rapid-fire
Define precision, recall, and F1.
Show answer
Precision is the share of predicted positives that are actually positive; recall is the share of actual positives you caught. F1 is their harmonic mean, which is high only when both are high. Which one matters more depends on the cost of a false alarm versus a miss. Full lesson →
⚡ Rapid-fire
What is a regression gate?
Show answer
An automated eval that runs in CI on every change to prompts, models, or indexes, and blocks the release if quality drops below an agreed bar. It stops one improvement from silently breaking cases that used to work. Full lesson →
⚡ Rapid-fire
What's the difference between direct and indirect prompt injection?
Show answer
Direct injection is the user typing instructions to override the system prompt. Indirect injection hides instructions inside content the model reads — a web page, an email, a retrieved document — so the attacker never talks to the model directly. Indirect is more dangerous for agents, because they read untrusted content and can take actions. Full lesson →
⚡ Rapid-fire
Can you fully prevent prompt injection?
Show answer
No — there's no complete fix today, because the model processes instructions and data in the same channel. You reduce the risk with layers: separate trusted from untrusted content, give tools least privilege, validate tool arguments outside the model, require confirmation for sensitive actions, and filter outputs. Full lesson →
⚡ Rapid-fire
What are guardrails?
Show answer
Checks wrapped around the model: on the input (blocking injection attempts or off-topic requests), on the output (PII, toxicity, policy violations, ungrounded claims), and on actions (validating tool calls). They can be rules, classifiers, or frameworks like NeMo Guardrails, and they work best layered rather than as a single check. Full lesson →
6 · Inference, serving & cost professional
⚡ Rapid-fire
What is the KV cache?
Show answer
During generation, the model stores the attention keys and values for every token it has already processed, so each new token only needs computing once instead of reprocessing the whole sequence. It makes generation much faster, but its memory grows with sequence length and batch size, which often limits how many requests a GPU can serve at once. Full lesson →
⚡ Rapid-fire
What does PagedAttention solve?
Show answer
Allocating one contiguous block of KV-cache memory per request wastes a lot of memory through fragmentation and over-reservation. PagedAttention, used in vLLM, stores the cache in fixed-size pages like operating-system virtual memory, so memory is allocated as needed and can be shared — fitting many more concurrent requests on the same GPU. Full lesson →
⚡ Rapid-fire
What is prompt caching, and how do you get hits?
Show answer
The provider reuses the processed form of a repeated prompt prefix, so cached input tokens are much cheaper and the response starts sooner. It only matches an exact prefix, so put stable content — system prompt, tools, reference documents — first and anything that changes per request after it. Full lesson →
⚡ Rapid-fire
What is continuous batching?
Show answer
Static batching waits for every request in a batch to finish before starting new ones, leaving the GPU idle while short requests wait on long ones. Continuous batching adds and removes requests at each generation step, keeping the GPU busy and greatly increasing throughput. Full lesson →
⚡ Rapid-fire
What does quantization trade off?
Show answer
Quantization stores model weights (and sometimes activations) at lower precision — for example 8-bit or 4-bit instead of 16-bit. It cuts memory and often speeds up inference, at the cost of some accuracy, which is usually small at 8-bit and needs measuring on your own task at 4-bit. Full lesson →
⚡ Rapid-fire
How does speculative decoding speed things up?
Show answer
A small, fast draft model proposes several tokens ahead, and the large model checks them all in a single forward pass, keeping the ones it agrees with. The output is the same as the large model would have produced on its own, but faster whenever the draft guesses well. Full lesson →
⚡ Rapid-fire
How do you estimate the cost of a request?
Show answer
Input tokens times the input price, plus output tokens times the output price, which is usually several times higher per token; cached input is billed at a discount. Then multiply by calls per task and tasks per month — agents and retries multiply calls quickly, so cost per completed task matters more than cost per call. Full lesson →
7 · Systems & reliability tech-lead
⚡ Rapid-fire
At-least-once versus exactly-once delivery?
Show answer
At-least-once means a message is retried until it's acknowledged, so it may arrive more than once; that's what most queues and webhooks give you. True exactly-once delivery isn't achievable across a network, so you get exactly-once effects by combining at-least-once delivery with idempotent processing. Full lesson →
⚡ Rapid-fire
What does idempotent mean, and how do you implement it?
Show answer
An operation is idempotent if doing it twice has the same effect as doing it once. You usually implement it with an idempotency key — a unique ID per request or event, stored under a unique constraint — so a repeat is detected and becomes a no-op. Full lesson →
⚡ Rapid-fire
Explain the CAP theorem in two sentences.
Show answer
When a network partition separates nodes of a distributed system, you have to choose between consistency (every read sees the latest write) and availability (every request gets a response). Partitions can't be ruled out, so the real choice is what the system does when one happens. Full lesson →
⚡ Rapid-fire
Why track p99 latency, not just the average?
Show answer
The average hides the slow tail, and the tail is what frustrated users actually experience. p99 is the latency that 99% of requests beat; when one page load makes many backend calls, most users hit at least one slow call, so the tail drives perceived performance. Full lesson →
⚡ Rapid-fire
How should a client handle rate limits?
Show answer
Retry on 429 and transient errors with exponential backoff plus random jitter, honour any retry-after header, and cap the number of attempts. The jitter matters: without it, many clients retry at the same moment and cause a thundering herd that keeps the service overloaded. Full lesson →
⚡ Rapid-fire
What is a circuit breaker?
Show answer
A wrapper around calls to a dependency that, after repeated failures, stops calling it for a cooldown period and fails fast or uses a fallback. It protects your service from piling up slow calls to something that's down, and it gives the dependency room to recover. Full lesson →
⚡ Rapid-fire
What signals would you monitor on an LLM feature in production?
Show answer
Latency (p50 and p99), errors and refusals, cost and tokens per request, and quality — measured by sampled LLM-judge scores and user feedback — all broken down by feature, tenant, and model version. Track drift over time too, because quality can decline slowly without any error firing. Full lesson →
🪜 Practice ladder beginner → industry
- Beginner: go through §2 with every answer closed, and answer each question out loud in two sentences.
- Easy: time yourself on all forty-two: 30 seconds per question. Mark every one you went over time on.
- Core: for each marked question, read the linked lesson section, then try the question again the next day.
- Stretch: have a peer fire fifteen random questions from this page at you in ten minutes, with no answers open.
- Hard: for five questions, add the follow-up an interviewer would ask next, and answer that too.
- Industry: explain three of these — say prompt caching, idempotency, and prompt injection — to a non-technical friend in one sentence each, the way you'd explain them to a client's executive.
✓ Checkpoint — you can move on when you can…
- Answer at least 37 of the 42 questions correctly with every answer closed.
- Keep each answer to two or three sentences and stop cleanly.
- Respond to a question you don't know by reasoning out loud, without bluffing.
- Name the lesson to revisit for any topic you're still shaky on.
Knowledge check check yourself
✓ Knowledge check
In a rapid-fire round, what does a strong answer look like, and what should you do when you don't know the answer?
Show answer
A strong answer is two or three sentences: a one-sentence definition, a sentence on why it matters or what it trades off, and then you stop and let the interviewer follow up. When you don't know, say so and reason out loud from what you do know ("I'm not sure, but here's how I'd reason about it"). Bluffing is easy to spot and costs more than the gap itself.
✓ Knowledge check
Why is exactly-once delivery achieved in practice through idempotency, rather than by the network or queue?
Show answer
Across a network you can't tell whether a lost acknowledgement means the message wasn't processed or was processed and the reply was lost, so reliable systems retry and deliver at least once — which means duplicates happen. Making processing idempotent, typically with a unique idempotency key, turns those duplicates into no-ops, giving exactly-once effects.
© 2026 studybydoing.in · AI Engineering: Zero to Production · All rights reserved. · About · Privacy Policy · Terms · Contact
Educational content, provided as-is and without warranty. Code samples are examples — review, test, and adapt them before using in production. See the Terms of Use & Disclaimer. Use at your own risk.