Product questions are knowledge questions

Customers ask whether a product fits, how it is used, what it contains, and when it is appropriate. Product pages may not contain every useful detail, while a language model may fill gaps with plausible guesses. The remedy is a maintained merchant knowledge base with clear ownership.

Write knowledge for retrieval

Use short entries with one question or fact per section. Include names, variants, dimensions, materials, care, compatibility, exclusions, links, and a last-reviewed date. Keep claims specific enough that a reviewer can compare the answer to the source quickly.

Ground before generating

Retrieve relevant entries first and require the draft to stay within them. If the source does not answer the question, the assistant should say it needs a human or ask a focused follow-up. Confidence should come from evidence, not tone.

Test uncomfortable cases

Include discontinued products, similar variants, missing attributes, contradictory entries, and questions that combine two products. Evaluation should check that the assistant avoids inventing facts and correctly signals uncertainty.

Keep the loop grounded.

Want to explore a support workflow with your own evidence and policies?

Start a pilot conversation