GenAI and LLM interview questions

Late chunking embeds a long passage with full context then pools token vectors into chunk embeddings so boundaries lose less meaning.

You store summaries of sections or documents, retrieve at the coarse level, then drill into child chunks.

SPLADE-style models expand queries and documents into sparse vocabulary-weighted vectors that still capture some synonyms.

A classifier or small LLM decides keyword-heavy versus semantic versus SQL versus web.

LTR or a cross-encoder consumes features such as BM25, vector score, recency, and click priors to reorder the union list.

Use namespaces or mandatory metadata filters on tenant ID and enforce them in application code, not only in prompts.

PQ compresses vectors into codebooks so RAM drops and distance is approximated.

Version the embedding model in metadata, dual-write or blue-green a new index, then atomically switch the query path.

Define maximum lag from source change to searchable vector, with alerts on ingest lag.

SFT clones demonstration answers via supervised loss.

A few thousand clean, diverse, on-policy examples usually beat a noisy dump of tickets.

QLoRA freezes a 4-bit quantized base model and trains LoRA adapters in higher precision.

If facts churn, require citations, or are tenant-specific, retrieve.

Validate schemas, timeout tools, retry idempotent reads, and never retry non-idempotent writes blindly.

JSON mode asks the model to emit JSON but may still be invalid or off-schema.

Every app reinvented ad-hoc tool schemas and auth.

Avoid it when you need a tiny, auditable control flow, custom streaming, or you are fighting abstraction leaks.

Trace each span: rewrite, retrieve, rerank, prompt, tokens, and tool calls.

Timeouts, partial streams, tool exceptions, and retry storms that duplicate side effects.

Fix temperature and seeds where possible, average multiple samples, and use paired tests on the same items.