Dataford
Interview QuestionsInterview GuidesExperiencesMock InterviewsPricing
Get started

Evaluate a Grounded Support Assistant

MediumGenerative AI & LLMs00:00
Practice interviewer
In session
5 left
00:00

Your question is Evaluate a Grounded Support Assistant. Take a moment with it on the right.

Talk me through your thinking if you like. When you're confident, submit your answer and I'll grade it like a real screen (7/10 or better passes).

You need to log in / sign up to chat or submit.

Problem

Scenario

You are responsible for evaluating an LLM assistant that drafts responses for operations and support teams using internal policy documents, underwriting guidelines, and customer account notes. The assistant is already live in a limited rollout and handles several thousand queries per day, but stakeholders do not trust the current quality signals because they rely mostly on thumbs-up feedback. You need an evaluation framework that measures answer quality, groundedness, and safety before the team expands usage. The system may answer directly, ask a clarifying question, or refuse when the source material is insufficient.

Constraints

  • p95 end-to-end latency for the assistant must stay under 2,500ms
  • Evaluation should support a cost ceiling of $12K/month at current traffic
  • Hallucinated factual claims must stay below 2% on a reviewed golden set
  • Prompt injection from retrieved documents or user messages must be treated as a real production risk
  • Responses must not expose sensitive customer or financial data beyond the user’s access scope

Available Resources

  • Historical prompts, retrieved context, model outputs, and agent traces from the pilot
  • Internal policy and account-document corpus with citation metadata
  • A GPT-4-class model and a smaller low-cost model for automated grading
  • 20 hours per week of expert reviewer time for labeling and adjudication

Question

How would you design a practical evaluation strategy for this LLM system so the team can measure response quality reliably and decide whether it is safe to scale? Explain how you would balance offline and online evaluation, define success metrics, and account for hallucination, prompt injection, latency, and cost in the final approach.