A field table is more clear than a long prompt
Before handing over scattered product information to AI, the product number, name, specifications, packaging, price conditions and image sources must be agreed upon. Which fields must come from the customer and which are allowed to be collated from existing text should be stated separately. When data is missing, the output is pending and no seemingly reasonable supplements are generated.
Check the handling method with conflict samples
Prepare samples with the same name but different specifications, old packaging and new packaging, inconsistent pictures and tables, etc., and observe whether the system marks conflicts. Each output field retains the location of the original data so that reviewers can review it quickly; don’t just use the smoothness of the copy as the acceptance criterion.
Manual priority is given to checking the amount, unit of measurement, applicable conditions and content related to qualifications. The determined product fact table is then used for expression in other channels, and the versions are synchronized when changes are made to avoid conflicting information on different pages. Dify's data collection and retrieval instructions provide technical background, which does not mean that automation itself can verify product facts.
Delivery and Maintenance
Deliver a unified catalog, missing field table, conflict handling record and update instructions to explain what content has been confirmed by the customer. Keep spot checks and stopping mechanisms before batch output. Extended reading: Product Catalog Cleaning Project.