CrackML by @ml.with.umang
Interview questions / GenAI & LLM
GenAI & LLM interview question

For a Q&A bot, when would you add RAG instead of relying on a parametric model alone?

For a Q&A bot, when would you add RAG instead of relying on a parametric model alone?

mediumsystem designEvidence 41/1001 source reportGoogle

The 60-second answer

Define the exact task boundary first: what the model should know, retrieve, generate, abstain on, and how success will be measured. Explain the relevant model or retrieval mechanism, including tokenization/embeddings, context construction, attention or decoding behavior, and the main failure modes.

Build the answer in this order

1
Define the mechanism

Define the exact task boundary first: what the model should know, retrieve, generate, abstain on, and how success will be measured.

2
Explain the architecture

Explain the relevant model or retrieval mechanism, including tokenization/embeddings, context construction, attention or decoding behavior, and the main failure modes.

3
Compare trade-offs

Evaluate components separately where possible: retrieval quality, groundedness, factuality, task success, latency, and cost; include representative slices and adversarial cases.

4
Close with serving + evaluation

Cover production controls such as access-aware retrieval, fallback/abstention, prompt/model versioning, monitoring, and rollback.

A useful interview mental model

This is the shape of a strong answer—not a script to memorize.

01Requirements
02Data
03Model / Retrieval
04Serving
05Monitor

Senior-level signal

  • Separate model capability from system reliability; a stronger model does not compensate for weak evidence, permissions, or evaluation design.
  • Discuss how you would detect silent regressions such as hallucination, evaluator bias, retrieval drift, or cost/latency creep.

What the interviewer is really testing

Understanding beyond prompting: architecture, retrieval, evaluation, inference, safety, latency, cost, and failure recovery.

Likely follow-up questions

What changes at 10× traffic or data volume?
Which failure mode would you monitor first in production?
How would you evaluate this offline and online before rollout?

Common weak-answer patterns

  • Jumping to a model before defining the product contract.
  • Listing components without bottlenecks, metrics, or failure handling.
  • Ignoring data quality, serving latency, monitoring, and iteration.