Knowledge Base Retrieval for Kindergarten Safety and Emergency Drills: Compare Documents and Locate Differences

Kindergarten Safety: review retrieval for document checks. Check source versions, input mappings and the values received by downstream nodes.

What this category's data looks like

The materials involved in kindergarten safety protection and emergency drills include hazard ledgers, rectification notices, drill plans, sign-in sheets, access control and school bus inspection records, and responsible person signature pages. These materials come from daily campus inspections, drill organization, and logistics security, and are updated weekly or per occurrence. Fields include hazard items, rectification deadlines, number of drill participants, sign-ins, inspection items, and responsible person signatures. When entering the knowledge base retrieval service, hazard ledgers and rectification notices serve as reference materials, drill plans and sign-in sheets serve as process materials, and access control and school bus inspection records serve as special materials. Each participates in recall in fragment form, for subsequent nodes or manual comparison of whether names, quantities, and dates are consistent.

Configuration and implementation checks

This step is executed by the knowledge base retrieval service (retrieval node and search test), outputting candidate fragments and similarity scores (after pattern, merge, rerank, filter, and limit trimming). The "discrepancy list" is completed by subsequent nodes or manually.

ObjectCheckVerification action
Hazard ledger and rectification noticeWhether hazard items and rectification deadlines are consistent, and whether hazards are closed-loopIn the search test, switch searchMode to hybrid, see whether candidates include both semantic and full-text results, and record the number of candidates and scores
Drill plan and sign-in sheetWhether the number of drill participants and sign-ins are consistent, and whether drill records match the planRetrieve once each with synonymous questions, compare candidates and scores, and observe changes in the order of the top few results before and after enabling usingReRank
Access control and school bus inspection recordsWhether inspection items are missing and whether responsible person signatures are completeCheck whether returned candidates cover inspection items and signature pages, and verify the key original text fragments retained after similarity filtering
Responsible person signature pageWhether signatures match the corresponding hazard or drill recordsRetrieve by combining object name and date, record the number of candidates and scores, and confirm that signature fragments and record fragments are hit at the same time

Three common errors

  • Phenomenon: The hazard ledger shows rectification completed, but the rectification notice has no closed-loop record. Cause: The two materials were ingested separately, and recall hit only one of them, without cross-material comparison for the same object.
  • Phenomenon: The number of drill sign-ins does not match the planned number in the plan. Cause: The similarity between the plan and sign-in sheet fragments was close, and after reranking the order of the top few results changed, and manual review did not check them one by one.
  • Phenomenon: Access control and school bus inspection records have missing inspection items. Cause: Inspection table fragments were trimmed by the reference limit, key inspection items did not enter the candidates, and it is necessary to verify the tokens of the referenced text and separately record the number of array items.

How to verify the setup

Acceptance of this feature: test once each with synonymously rewritten questions, compare candidates and scores, then toggle reranking on and off to observe the ordering of the top few results; at the same time switch searchMode, record the number of candidates and scores, and observe the impact of limit and similarity on the returned results.

Business test: use one real record to find fields inconsistent with the ledger, hand it to the reviewer for confirmation, and write it back.

  • In the knowledge base search test, retrieve the same object separately in semantic, full-text, and hybrid modes, record the number of candidates and scores, and compare the ordering of the top few results.
  • Retrieve once each with reranking enabled and disabled, verify whether candidates include both semantic and full-text results, and confirm that the tokens of the referenced text and the number of array items are recorded separately.