Key Concepts
- 1In commerce, a hallucination — an invented product, price, policy, or availability — is not a quirk but a trust and potentially legal failure, so its rate is a first-class product metric
- 2Grounding via RAG reduces but doesn't eliminate hallucination; the agent can still misstate a grounded fact or fabricate when context is thin
- 3Measuring hallucination requires ground truth: comparing agent claims against the authoritative catalog, pricing, and policy systems, often via sampling and human review
- 4Not all hallucinations are equal: a wrong product description is minor; a fabricated price, promise, or policy the user relies on is severe
- 5Guardrails help: constraining the agent to only state facts from authoritative sources and refusing to answer when ungrounded lowers the rate
- 6Hallucination interacts with helpfulness: an agent tuned to always answer hallucinates more; one tuned to say 'I don't know' is safer but can feel less capable — a real tradeoff
- 7Track hallucination alongside conversion and satisfaction, because reducing it sometimes costs perceived helpfulness, and the balance is a product decision
Why It Matters for PMs
Trust and legal stakes: A fabricated price or policy the user acts on is a serious failure. Treating hallucination rate as a tracked metric forces accountability for it.
Grounding isn't a silver bullet: RAG reduces but doesn't eliminate hallucination. PMs must measure and manage the residual rate, not assume it's solved.
Helpfulness tradeoff: Reducing hallucination can reduce perceived capability. Balancing the two is a product decision that requires measuring both sides.
Test Your Understanding
Question 1 (Multiple Choice)
Why is hallucination a first-class metric in commerce specifically?
Question 2 (Multiple Choice)
Does RAG grounding eliminate hallucination?
Question 3 (Multiple Choice)
What tradeoff comes with tuning an agent to hallucinate less?
Question 4 (Short Answer)
How would you measure an agent's hallucination rate in a commerce context?
Think about your answer before revealing the explanation.
Question 5 (Short Answer)
How would you weigh reducing hallucination against preserving the agent's perceived helpfulness?
Think about your answer before revealing the explanation.