Objective
You should leave with a framework for "cite vs guess" decisions in product flows.
Hallucination basics
A model can produce confident but wrong statements when:
- input lacks needed facts,
- retrieval misses key documents,
- prompt rewards fluency over evidence.
Grounding strategy
- Retrieve relevant documents.
- Pass excerpts with identifiers.
- Instruct the model to answer only from sources.
- Require explicit citation or abstention.
Worked example (online store)
Customer asks: "Can I return final-sale shoes after 30 days?"
Without grounding, model may fabricate a policy exception. With RAG, the system retrieves current return policy and asks model to cite the clause.
If no clause is found, response should say "I don't have enough evidence" rather than inventing.
Product decision
Reliability is not free. Retrieval pipelines add latency and cost, but reduce risk and support burden.
Three takeaways
- Fluent text is not proof.
- Citations turn generation into auditable behavior.
- Abstention is often the safest response.