GenAI and LLM interview questions

The model returns a function name and JSON arguments instead of a user-facing answer.

Schemas tell the model required fields, types, and enums so arguments are usable.

The model may emit several function calls in one turn, for example weather in two cities.

Live APIs and databases supply ground truth the weights never had.

Offline evals run a frozen set of cases in CI.

Faithfulness asks whether claims are supported by retrieved context.

They reward n-gram overlap with a reference, but many correct phrasings exist.

Keep a labeled suite, run it on every prompt diff, and fail the build on score drops.

Jailbreaks target the model's safety policy, usually in the user turn.

Use a clear system policy, output classifiers, and refusal evals.

They scan the user input and the model output for hate, self-harm, sexual content, or PII.

The model may regurgitate copyrighted or private strings it saw in pretraining or fine-tuning.

They are a separate model or rules engine that scores text independently of the chat model.

Tighten prompts, cap history, retrieve fewer better chunks, and cache embeddings and repeated prefixes.

Larger models and longer chain-of-thought raise quality and delay the first token.

Offline scoring, embeddings for a corpus, and nightly reports can use batch APIs at lower price.

Use exponential backoff with jitter, queue work, and fall back to a smaller model or cached answer.

Chunk and retrieve, map-reduce summarize, or use a long-context model for the last mile only.

Models often use the start and end of a long prompt more than the middle.

They reduce aggressive chunking and can attend over a whole report.