What this category's data looks like
This page addresses the kindergarten health care and meal management scenario. The materials come from five types of ledgers: quantified recipes, food ingredient purchase orders, sample retention registration, disinfection records, and absence cause registration. They are updated on a dual monthly and weekly cadence, and the fields include ingredient name and specification, sample retention time and batch, disinfection time period, and absence cause classification. Ingredient name and specification are the primary key of the terminology table, determining whether the same ingredient is written consistently in the recipe, purchasing, and sample retention locations; sample retention time and batch are the traceability anchors, used to align documents with sample retention batches; disinfection time period is a time-period-type field, covering nodes such as morning check, before meals, and departure; absence cause classification is a controlled vocabulary, and when crossing languages it is necessary to provide translated names that can refer back to the original text. When the materials enter the multi-turn dialogue and prompt stages, they respectively serve as terminology candidates, batch evidence, time-period constraints, and classification mappings, based on which the model generates a unified terminology table and a comparison draft.
Configuration and implementation checks
This part is executed by the memory turn count and prompt configuration in the application settings, outputting the referenced historical message range and answer constraints. The unified terminology table and comparison draft are then completed by the model, while the external entry point and authorization are handled by the publishing service and file permissions.
| Object | Check | Verification action |
|---|---|---|
| Ingredient name and specification | The same ingredient is written consistently in the recipe and purchasing categories, and the translated name can refer back to the original text | Check the memory turn count value in the application settings, run a trial with two consecutive question turns, and compare the ingredient wording in messages |
| Sample retention time and batch | The document and sample retention batch match, and the batch number is consistent throughout the comparison draft | Extract three sample retention records and check whether the batch wording is consistent within the conversation corresponding to chatId |
| Disinfection time period | The disinfection records have no missing time periods, and the translated time period names correspond one-to-one with the original nodes | Spot-check the time period fields at the three nodes of morning check, before meals, and departure, and record the time period values in variables |
| Absence cause classification | The classification vocabulary is consistent throughout, and the translated names can refer back to the original cause descriptions | Extract three causes and check whether their classification wording is unified in the comparison draft |
Three common errors
- Phenomenon: The document and sample retention batch do not match. Cause: Sample retention registration and purchase orders are maintained separately, batch numbers are not checked against the same source, and cross-language translated names are each transcribed differently.
- Phenomenon: The recipe and purchasing categories are inconsistent. Cause: Ingredient name and specification have not entered the terminology table, the recipe uses common names while purchasing uses specification names, and the translated names cannot refer back to each other.
- Phenomenon: Disinfection records are missing time periods. Cause: Time period fields are filled in separately by node, the turn count is too small to use prior context constraints, and the model omits nodes when completing them.
How to verify the setup
Acceptance of this feature: Ask two consecutive follow-up questions and see whether the second question inherits the reference from the first question; check whether the historical range and answer constraints in messages match.
Business test: Extract three terms and check whether their wording is consistent throughout, and confirm that the translated names can refer back to the original text.
- Check the current value of the memory turn count in the application settings and match it with the prior context range required by the question.
- Run a trial with two consecutive question turns, compare the usage records of the two runs, and confirm that consumption changes accordingly after adjusting the turn count.