AI Data Engineering Interview Questions

Retrieval-Augmented Generation retrieves relevant documents at query time and passes them to the LLM as context. It keeps knowledge fresh without retraining, gives citations, respects permissions per user, and is cheaper to update. Fine-tuning changes behaviour/style, not up-to-date knowledge.

Balance context completeness against precision: 300–800 tokens with 10–20% overlap as a start; structure-aware splitting (headings, code blocks, tables). Then tune on a retrieval eval set (recall@k), not intuition.

Dense embeddings capture semantics but miss exact tokens (IDs, error codes, names); BM25 nails exact terms. Combining (e.g. reciprocal rank fusion) improves recall across query types.

Ingest ACLs as chunk metadata, sync permission changes quickly, and pre-filter retrieval by the user’s principals. Never rely on the LLM to hide content; audit what was retrieved for whom.

Incremental ingestion with content hashing (skip unchanged), deterministic chunk ids for idempotent upserts, delete propagation, a source-of-truth chunks table (Delta) from which the index is synced/rebuilt, and freshness monitoring.

Vectors from different models aren’t comparable, so re-embed everything into a new index version in parallel, evaluate, then switch an alias (blue/green). Track model version per vector.

HNSW: graph-based, high recall and speed, memory-hungry, incremental inserts. IVF(-PQ): cluster-based, less memory (with product quantisation), needs training, typically lower recall. Quantisation (int8/binary) trades a little recall for 4–32× memory savings.

Retrieval: recall@k / MRR / nDCG on a golden set of questions with relevant chunks. Generation: faithfulness/groundedness and answer relevance (LLM-as-judge calibrated with human labels). System: latency, cost, refusal rate. Business: user feedback, deflection.

request/trace id, app, user (pseudonymised), model + version, prompt version, input/output/cached tokens, cost, latency, retrieved chunk ids + scores, tool calls, status/errors, user feedback, eval scores. This becomes a fact table for cost, quality and regression analysis.

Versioned judge prompts/models, calibration against human labels (agreement metrics), checks for known biases (length, position, self-preference), pairwise comparisons for A/B decisions, and spot checks.

Treat prompts/configs as code; run an offline eval suite in CI on golden datasets; canary/shadow in production comparing metrics with confidence intervals; rollback on regression.

A system to define, compute, store and serve ML features consistently: offline store (history for training, point-in-time joins), online store (low-latency latest values for inference), registry (definitions, owners, lineage).

Training examples must use feature values as known at the label’s timestamp. Joining current values leaks future information, inflating offline metrics and failing in production. Use AS-OF joins or logged features.

Different feature computations (or data) at training vs inference time causing different values. Prevent with single feature definitions materialised to both stores, feature logging at serving time, and distribution monitoring.

Inventory and classify SQL; deterministic transpiler first; LLM agent for the long tail with retrieval of similar solved examples; compile checks; result-set equivalence validation against the legacy system; repair loop with error feedback; human review queue for low confidence; every fix feeds the knowledge base; metrics on auto-validation rate and human hours per object.

Least-privilege identities (agent acts as the user), governed tables with good metadata (descriptions, semantic layer) for grounding, query cost/time limits, read-only by default, audit logging of generated queries, validation of results, and evaluation on a benchmark of questions.

Cache (prompt/semantic caching), batch requests, route to smaller models when possible, trim context (better retrieval, fewer chunks), cap agent steps/tokens, sample expensive evaluations, monitor cost per request per app with budgets and alerts.