How Would You Prevent Hallucinations in an LLM-Based System?

You're interviewing for an AI-oriented Software Engineer position at a company that builds products on large language models. In the screening interview, the interviewer poses this question: "How would you prevent — or at least substantially reduce — hallucinations in an LLM-based system?" Explain the sources of hallucination and the specific methods you would use across the stack (training, inference, retrieval/grounding, verification, product/UX, and evaluation) so the system does not confidently assert things that are false or unsupported.

Hint — Where to start: Don't reach for one fix immediately. Begin by defining what "hallucination" means for this product (unsupported claims vs. factually false claims vs. invented citations), then group mitigations by where in the lifecycle they operate: before generation (data/training), during generation (grounding + decoding), and after generation (verification + UX). Hint — The highest-leverage technique: In most production systems, the largest single gain comes from grounding: retrieval-augmented generation (RAG) or tool use that places authoritative context into the prompt, along with instructions (and training) that push the model to answer only from that context and otherwise abstain. Then consider how you would verify the output against the retrieved sources.

Constraints & Assumptions

• Assume a production LLM-based application (for example, a question-answering assistant or agent), not a research prototype.
• You may or may not control model pretraining; assume you can fine-tune, prompt, and build surrounding infrastructure.
• Latency and cost matter: mitigations that multiply inference cost require justification.
• Interpret "prevent" honestly — hallucination can be reduced and contained, but not eliminated with certainty.

Clarifying Questions to Ask

• What kind of product is this — open-domain chat, domain-specific Q&A over private data, code generation, or an autonomous agent? The dominant failure mode is different in each case.
• How expensive is a hallucination here? Is this a casual assistant, or a high-stakes domain (medical, legal, financial) where an incorrect answer causes real harm?
• Do we control the model (can fine-tune / RLHF), or are we using a third-party API where only prompting and system-level defenses are available?
• Is there an authoritative source of truth we can ground against (documents, database, APIs), or is the model expected to answer from parametric knowledge?
• What latency and cost budget do we have per request (this affects whether multi-pass verification or self-consistency is feasible)?

What a Strong Answer Covers

Follow-up Questions

• Your RAG system still hallucinates: the answer contradicts the retrieved passages. What are the likely causes, and how would you fix each one?
• How would you build an automated evaluation pipeline to measure hallucination rate on every model or prompt change, and what would you use as ground truth?
• When is it better for the model to say "I don't know," and how do you actually train or prompt a model to abstain without making it uselessly evasive?
• How do these mitigations change for an agent that takes actions (calls APIs, writes code) rather than one that only produces text? Overview: This question tests a candidate's grasp of hallucination in large language models, including underlying causes, detection, and system-level mitigation approaches, as well as their ability to reason about model reliability and failure modes.