What this category's data looks like
This category covers three types of materials: exam administration manuals, examination room record sheets, and disciplinary violation fact records. They come from the exam administration department and test site receipts, and their update cadence changes with exam batches and rule revisions. The exam administration manual carries the process and discretionary standards, the examination room record sheet carries the room status and timeline for that session, and the disciplinary violation fact record carries the parties involved, the conduct, and the signed receipt. When these three enter knowledge base retrieval and recall, the exam administration manual is mostly semantically recalled as a rule basis, while the examination room record sheet and the disciplinary violation fact record are mostly recalled as factual fragments via full-text or hybrid recall. Field characteristics are concentrated in four categories: name, quantity, date, and signature. When the same record does not match across two materials or versions are not synchronized, retrieval candidates will present multiple standards at the same time. It is necessary to first distinguish source and version at the candidate and score level, and then hand off to subsequent nodes or manual work to compile the version ledger and change records.
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 mode, merge, rerank, filter, and limit truncation). The version ledger and change records are then completed by subsequent nodes or manual work.
| Object | Check | Verification action |
|---|---|---|
| Exam administration manual | Whether the name, quantity, and date of the same record are consistent across the three types of materials, and whether the current version is unique | In the search test, switch searchMode, run one retrieval each with synonymous questions, record the number of candidates and similarity, and check the version and date to which the hit fragments belong |
| Examination room record sheet | Whether the name, quantity, date, signature, and receipt of the disciplinary violation fact record and the examination room record sheet match | Enable usingReRank to compare the order of the top few results, check whether candidates include both semantic and full-text results, and verify the signature and receipt fields one by one |
| Disciplinary violation fact record | Whether the old version is traceable, whether references are synchronized, and whether fragments missing a signature or receipt are still recalled | Record the number of array items by limit, separately check the tokens of the reference text, and confirm that key fragments of the original text are still retained after truncation |
| Retrieval node | Whether candidates and scores correspond to the material version, and whether reference limit truncation only affects the number of items | After changing one piece of material content, rerun retrieval, compare the number of candidates, scores, and reference text, and confirm that the version number and references are updated in sync |
Three common errors
- Symptom: The name, quantity, and date of the same disciplinary violation record do not match between the examination room record sheet and the disciplinary violation fact record. Cause: The two materials are maintained separately, and fields were not checked against the version ledger.
- Symptom: The exam administration manual has been revised, but retrieval still recalls old-version fragments, and references are inconsistent with the current standards. Cause: Versions are not synchronized, and the old version was not marked or removed from candidates.
- Symptom: A disciplinary violation fact record missing a signature or receipt still enters references as a candidate. Cause: Material completeness was not verified before retrieval, and filtering was performed only by similarity.
How to verify the setup
Acceptance for this feature: test once each with synonymously rewritten questions, compare candidates and similarity scores, then toggle usingReRank to observe the ranking of the top few results.
Business test: change one piece of content, see whether the version number and references are updated in sync, and record the number of candidates and the tokens of the reference text.
- In the search test, switch
searchMode, separately record the number of candidates and scores for semantic, full-text, and hybrid modes, and check the material and version to which the hit fragments belong. - Record the number of array items by
limit, separately check the tokens of the reference text, and confirm that key fragments are still retained after similarity filtering and reference limit truncation.