Question 1

Context windows in 2025-26: how do you use 128k-1M tokens without drowning?

Accepted Answer

**Target Audience:** Engineers who have a 1M-token model available and need to decide when to use the window and when to keep retrieving.
**Difficulty:** Easy
**Verified with:** Node 22.22.0 and Python 3.12.3 (cost arithmetic, chunk reordering and a needle-test harness with canned answers; no real model is called). Model window sizes, prices and published guidance are cited inline with the date checked (October 2026); these change often, so check the provider docs for the model you use.

> **How

Question 2

What is RAG and why use it instead of fine-tuning for knowledge?

Accepted Answer

**Target Audience:** Beginners meeting retrieval-augmented generation for the first time, and engineers who need to defend "RAG, not fine-tuning" for a knowledge feature in an interview.
**Difficulty:** Easy
**Verified with:** Node 22.22.0 and Python 3.12.3 (a keyword retriever plus deterministic stub models; no real model or vector database is called). Papers cited inline were checked in October 2026.

> **How to read this doc:** Every idea is explained in plain language first, then tagged with

Question 3

What are hallucinations and what causes them?

Accepted Answer

**Target Audience:** Anyone new to building with LLMs who has seen a model state something false with total confidence and wants to understand why, in plain terms.
**Difficulty:** Easy
**Verified with:** Node 22.22.0 and Python 3.12.3 (a counting toy, an expected-score calculation and a string-matching grounding check; no model is called). Paper and case facts are cited inline with the date they were checked (October 2026).

> **How to read this doc:** Every idea is explained in plain language f

Question 4

When do you use code-based graders versus model-based graders?

Accepted Answer

**Target Audience:** Engineers writing their first eval suite who need to decide, check by check, whether to write a function or ask a model, and candidates who should be able to defend that choice.
**Difficulty:** Medium
**Verified with:** Node 22 and Python 3.12 (five code graders and a stub judge over ten outputs, and a judge-versus-human validation with corrected pass rates). Judge verdicts are canned stand-ins for model calls; token prices in the cost line are assumptions. Vendor tooling is

Question 5

How do you reduce hallucinations in a production LLM application?

Accepted Answer

**Target Audience:** Engineers who own an LLM feature that real users rely on (support bots, internal assistants, document Q&A) and need to make it stop inventing things, measurably.
**Difficulty:** Medium
**Verified with:** Node 22.22.0 and Python 3.12.3 (the four-layer scoring run below). Case law, papers and provider features were checked via the sources cited inline, in October 2026.

> **How to read this doc:** Every idea is explained in plain language first, then tagged with `📌 Interview

Question 6

How would you design a production RAG pipeline end-to-end in 2025-26?

Accepted Answer

**Target Audience:** Engineers preparing for an AI or backend system design round who can explain RAG in one sentence and now need to design the whole thing: ingestion, retrieval, generation, numbers, operations and evaluation.
**Difficulty:** Hard
**Verified with:** Node 22.22.0 and Python 3.12.3 (the miniature pipeline and the sizing arithmetic below). Every input to the sizing is a stated assumption, and the prices in it are illustrative placeholders, not a vendor price list. Product facts ar

Question 7

Why are BLEU and ROUGE poor metrics for LLM outputs, and what replaced them?

Accepted Answer

**Target Audience:** Engineers who have seen BLEU or ROUGE in a dashboard or a paper and need to explain why they mislead for LLM features, and candidates expected to know the modern metric stack.
**Difficulty:** Hard
**Verified with:** Node 22 and Python 3.12 (BLEU, ROUGE-1 and ROUGE-L implemented from scratch, scored on hand-written answers next to a key-fact check, then against different references). Published results are cited by paper, checked October 2026.

> **How to read this doc:** Ever

Question 8

What are common RAG failure modes and how do you debug them?

Accepted Answer

**Target Audience:** Engineers who run or are about to run a RAG feature in production and need a systematic way to turn "the bot got it wrong" into a specific, fixable cause.
**Difficulty:** Hard
**Verified with:** Node 22.22.0 and Python 3.12.3 (the chunk-boundary demo and the triage harness below). The request logs in the triage are hand-written examples of each failure type, not data from a real system. Research findings are cited inline.

> **How to read this doc:** Every idea is explained

Question 9

Fine-tune vs RAG vs prompt engineering: how do you choose in 2025-26?

Accepted Answer

**Target Audience:** Engineers and tech leads who must choose how to adapt an LLM for a real product, and want a decision process they can defend with numbers rather than opinions.
**Difficulty:** Hard
**Verified with:** Node 22.22.0 and Python 3.12.3 (a freshness simulation, a cost model with **assumed** prices, and a rule-based decision table; no model is called). Provider and paper facts are cited inline with the date they were checked (October 2026).

> **How to read this doc:** Every idea i

AI ENGINEERINGDECODED.

Easy