What this category's data looks like
This page focuses on the kindergarten health care and meal management scenario. Materials include food procurement orders and specification descriptions, retention sample records and batch labels, disinfection period registration forms, absence cause classification tables, quantified recipes, and daily class attendance reports. These materials enter the knowledge base retrieval service in a mixed form of structured tables and text. Procurement orders and recipes play the role of category and specification comparison, retention sample records and batch labels play the role of time and batch tracing, disinfection registration forms play the role of period completeness verification, and absence cause classification tables play the role of attribution consistency. Materials are updated daily or by meal. Fields include ingredient name, specification, retention sample time, batch number, disinfection start and end times, and absence cause classification. The retrieval service is only responsible for recalling candidate snippets and similarity scores. The exception list and review conclusions are generated by subsequent nodes or manually based on this, and handed to the responsible person for handling.
Configuration and implementation checks
This step is executed by the knowledge base retrieval service (retrieval node and search test), outputting candidate snippets and similarity scores (after pattern, merge, rerank, filter, and limit trimming). The exception list and review conclusions are then completed by subsequent nodes or manually.
| Object | Check | Verification action |
|---|---|---|
| Food procurement orders and quantified recipes | Whether ingredient names and specifications are consistent, and whether recipes match procurement categories | In the search test, search separately by ingredient name and specification, switch searchMode to compare candidate snippets, and check whether the hit snippets cover procurement categories and recipe quantities |
| Retention sample records and batch labels | Whether retention sample time and batch correspond to the documents, and whether the documents match the retention sample batch | Search by retention sample time and batch number, view the number of candidates and scores, enable usingReRank to compare the ranking of the top few, and confirm whether batch label snippets are recalled |
| Disinfection period registration forms | Whether disinfection periods are complete and whether any period is missing | Search by disinfection period and date, view the number of returned candidates and scores, check whether snippets cover start and end times, record the array count separately, and verify the referenced text tokens |
| Absence cause classification tables | Whether absence cause classification is consistent with the original records | Search by absence cause classification, adjust similarity to observe candidate changes, and check whether hit snippets support the classification conclusion |
Three common errors
- Phenomenon: documents and retention sample batches do not match. Cause: retrieval candidates do not cover batch label snippets, or similarity filtering cuts off key snippets. First verify hits and associations.
- Phenomenon: recipes and procurement categories are inconsistent. Cause: semantic recall favors recipe text, and procurement order snippets are not merged into candidates. Switch retrieval modes for comparison.
- Phenomenon: disinfection records are missing periods. Cause: full-text recall does not hit start and end time snippets, and after reference limit trimming only the summary remains. Verify the completeness of referenced text.
How to verify the setup
This feature acceptance: test once each with synonym-rewritten questions, compare the number of candidates and scores, then toggle rerank to see the ranking of the top few, switch searchMode, and adjust similarity and limit to observe candidate changes.
Business test: create an exception record, see whether the source material can be traced, confirm that the exception is supported by original records and signed by the reviewer, and then hand it to the responsible person for handling.
- In the knowledge base search test, switch retrieval modes, compare the number of candidates and scores, record the array count separately, and verify the referenced text tokens.
- Compare the ranking of the top few before and after enabling rerank, check whether candidates include both semantic and full-text results, and confirm that key snippets are not trimmed.