What this category's data looks like
Graduate admissions re-examination and admission materials include re-examination plans, candidate score sheets, transfer lists, etc. They are published by the graduate schools and departments of admissions units. The update rhythm is concentrated in the re-examination season and transfer window. Fields involve candidate name, ID, score, date, and signature. When entering knowledge base retrieval and recall, the re-examination plan plays the role of rule basis, the candidate score sheet plays the role of numerical verification, the transfer list plays the role of list and status, and the version ledger and change records are used to mark the current version, trace old versions, and synchronize references. Material forms include tables, notices, and receipts. Retrieval needs to match fields and dates to avoid the same record having inconsistent names, quantities, or dates across two materials.
Configuration and implementation checks
This part 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 version ledger and change records are completed by subsequent nodes or manually.
| Object | Check | Verification action |
|---|---|---|
| Re-examination plan | Fields include re-examination date, score weight, and signature column; the same record must have consistent date and name across two materials | Switch searchMode in search test, test once each with synonymous questions, compare candidate count and score, and check whether date and signature match |
| Candidate score sheet | Fields include candidate ID, score, and date; quantity and name must not be inconsistent across two materials | Turn usingReRank on and off to compare the ranking of the first few items, check similarity score and reference text tokens, and record array count separately |
| Transfer list | Fields include candidate name, transfer status, and receipt; missing signature or receipt is considered inconsistent | Check whether candidates include both semantic and full-text results, look at the number of references and content completeness after limit trimming, and confirm old versions are traceable |
| Version ledger | Current version is unique, old versions are traceable, references are synchronized, and version number and references must be updated synchronously | After changing one piece of content, rerun retrieval to see whether candidates and scores change and whether reference text is updated synchronously |
Three common errors
- Symptom: The same candidate record has inconsistent name, quantity, or date between the score sheet and transfer list. Cause: The two materials have not aligned fields and dates according to the version ledger.
- Symptom: The re-examination plan has been updated, but retrieval still returns old clauses. Cause: The version was not synchronized, and references were not updated with the version number.
- Symptom: The transfer list lacks a signature or receipt, and the answer lacks key paragraphs. Cause: Although there are many candidates, the ones ranked first are irrelevant, and rerank and filtering were not performed.
How to verify the setup
Acceptance for this feature: test once each with synonymously rewritten questions, compare candidates and scores, then toggle rerank to see the ranking of the first few items; record candidate count, similarity score, and reference text tokens, record array count separately, and do not compare quoteMaxToken with the count.
Business test: change one piece of content and see whether the version number and references are updated synchronously; confirm the current version is unique, old versions are traceable, and references are synchronized; mark as inconsistent when signature or receipt is missing.
- In the knowledge base search test, switch
searchMode, test once each with synonymous questions, compare candidate count and score, checksimilarityscore and reference text tokens, and record array count separately. - Turn
usingReRankon and off to compare the ranking of the first few items, check the number of references and content completeness afterlimittrimming, and confirm candidates include both semantic and full-text results.