Generative AI for Data Engineers interview questions

A neural network trained on a lot of text to predict the next token, which lets it write, summarise, translate and reason.

Numeric vectors that represent meaning, so similar texts have similar vectors. Used for search and RAG.

Retrieval augmented generation: retrieve relevant chunks from your data and give them to the model, so answers are grounded and current.

A store that indexes embeddings for fast similarity search, like pgvector, Databricks Vector Search or Snowflake Cortex Search.

Split by structure (headings, paragraphs) into a few hundred tokens with some overlap, and keep metadata like source and section.

Measure retrieval (did the right chunks come back) and answer quality (correct, grounded, cites sources) on a test set.

The model returns a structured request to call a function you define, like run_sql, and your code executes it.

A model that plans and takes multiple steps with tools to reach a goal, checking results along the way.

Model Context Protocol, an open standard to connect AI applications to tools and data sources in a consistent way.

Extract text, prompt for a fixed JSON schema, validate types in code, and send failures to a review queue.

Read-only credentials, allow-listed schemas, query limits and timeouts, SQL validation before running, and logging.

Use smaller models where possible, cache repeated prompts, keep prompts short, batch work and set usage limits.

Mask or remove personal data before sending it, use approved providers and regions, and log access.

RAG adds knowledge at query time and is easier to update. Fine-tuning changes behaviour or style. Most data use cases start with RAG.

Snowflake Cortex functions, Databricks AI functions and Fabric Copilot let you call LLMs from SQL or notebooks.

Low temperature, strict schemas, validation, retries with error messages, and a fallback path.

Time to answer. Reduce with smaller models, shorter prompts, streaming responses and parallel calls.

Checks on inputs and outputs: blocked topics, PII filters, schema validation and rate limits.

Writing and reviewing SQL and PySpark, generating tests and documentation, explaining errors and refactoring code.

Log prompts, responses, latency, cost and errors, track quality on samples, and alert on failures or spikes.