GenAI Engineering Interview Prep

GenAI Engineering Interview Topics

• RAG vs Fine-tuning — RAG: retrieval at runtime, dynamic, no training; Fine-tuning: static knowledge baked into weights
• Embeddings — dense vectors capturing semantic meaning; cosine similarity measures relatedness
• Vector database — stores embeddings, enables fast approximate nearest-neighbour (ANN) search
• Hallucination — LLM generates confident but incorrect content; prevent with RAG, citations, low temp
• Agent vs chain — chain is a predefined sequence; agent decides dynamically which tools to use
• Tool use / function calling — LLM selects external functions with structured input schemas
• Context window — max tokens processed at once; manage with sliding history and summarisation
• Semantic cache — reuse responses for similar queries by embedding similarity threshold
• LLM-as-judge — use an LLM to rate output quality on a rubric (accuracy, clarity, hallucination)
• Chunking — split docs into pieces for indexing; size and overlap affect retrieval precision
• LoRA — train small adapter weights instead of full model; 100x fewer trainable parameters
• Guardrails — input validation + output checking to prevent harmful or off-topic LLM responses
• Streaming — deliver tokens as generated for better perceived performance
• Prompt caching — Anthropic caches repeated system prompts for up to 90% token cost reduction