Knowledge Base Retrieval for Student Discipline and Appeals: Find Supporting Evidence

Student Appeals: review retrieval for evidence lookup. Compare the source records, required inputs and downstream results before handing the task over.

What this category's data looks like

This page focuses on student disciplinary and appeal handling scenarios. The materials include three types: disciplinary fact materials, disciplinary decision documents, and appeal statements. Disciplinary fact materials record the course of conduct and evidence; disciplinary decision documents state the basis clauses and effective dates; appeal statements present demands and receipts. Materials are updated by semester batches, versions are identified by document number and date, and fields include name, student ID, reason, clause, date, and signature. When entering knowledge base retrieval, the three types of materials serve as candidate segment sources respectively. After question optimization, recall, merging, reranking, and filtering, they form citable content that can be traced back to the original text, for subsequent nodes or manual comparison of fields and dates.

Configuration and implementation checks

This step is executed by the knowledge base retrieval service (retrieval node and search test), outputting candidate segments and similarity scores (after mode, merging, reranking, filtering, and upper-limit trimming). "Basis items that can be traced back to the original text" are then completed by subsequent nodes or manually.

ObjectCheckVerification action
Disciplinary fact materialsName, quantity, and date must match the disciplinary decision document; signatures and receipts must be completeIn the search test, switch searchMode to hybrid, and check whether the candidate segments cover both the facts and the clauses of the decision document
Disciplinary decision documentThe version is current, clauses and effective dates are consistent; missing signatures or receipts must be markedCompare the ranking of the top few items before and after enabling usingReRank, and record the number of candidate items and the similarity score
Appeal statementThe demand and receipt date are consistent with the disciplinary decision document, and the same record name does not conflictAdjust the number of cited items by limit, and check the returned citation text tokens and array item count, recording them separately
Knowledge base retrieval serviceOutputs candidate segments and similarity scores, and does not output statistical tables or authorization resultsTest once each with synonymous questions, compare candidates and scores, and confirm that key segments of the original text are still retained after trimming

Three common errors

  • Phenomenon: The name and quantity of the same record do not match between the disciplinary fact materials and the disciplinary decision document. Cause: The two entries were not checked against the same document number and date, and the fields were not aligned.
  • Phenomenon: The version of the disciplinary decision document is not synchronized, and old clauses are still recalled. Cause: The update batch did not cover the old files, and retrieval still hits historical versions.
  • Phenomenon: The appeal statement lacks a signature or receipt, but the answer still gives basis items. Cause: The materials were entered into the knowledge base before being completed, and filtering and reranking did not intercept segments with missing items.

How to verify the setup

This feature acceptance: test once each with synonymously rewritten questions, compare candidates and similarity scores, then toggle usingReRank to see the ranking of the top few items, and use limit to adjust the number of items.

Business test: ask a basis question across materials, and see whether the answer gives a traceable source.

  • In the search test, switch searchMode, check whether the candidates include both semantic and full-text results, and record the number of candidate items and scores.
  • Check the tokens of the returned citation text and the array item count, record them separately, and confirm that key segments of the original text are still retained after trimming.