GenAI and LLM interview questions

Temperature rescales the logits before the token distribution is sampled.

Temperature zero makes decoding greedy: the highest-probability token is chosen at each step.

Top-p sampling keeps the smallest set of tokens whose probabilities sum to p, then samples from that set.

Fine-tuning continues training on your examples so weights change.

Choose RAG when knowledge must stay fresh, be cited, or remain outside the weights for privacy reasons.

Instruction tuning is supervised fine-tuning on prompts paired with desired responses.

LangChain is a framework for wiring prompts, models, retrievers, tools, and memory into applications.

A chain is a pipeline that passes data through prompt templates, models, parsers, and other steps.

A simple chain follows a fixed graph of steps.

Function calling lets the model request a structured tool invocation instead of only free text.

Downstream code needs fields it can parse, not a paragraph.

Evals are tests that score model or prompt quality on a dataset.

A golden dataset is a curated set of inputs with expected answers or grading notes.

LLM-as-a-judge uses another model to score an answer against a rubric.

Prompt injection is hostile text that tries to override your instructions, often by saying ignore previous rules.

Guardrails are checks that filter inputs and outputs for policy, PII, jailbreaks, and schema validity.

PII is personally identifiable information such as names, emails, and IDs.

RLHF is reinforcement learning from human feedback.

Most APIs bill per million tokens, often with different rates for input and output.

Generation is autoregressive and more expensive to serve than a single prompt encode.