What this category's data looks like
In the business of general higher-education admissions enrollment and filing, three types of materials—enrollment plans, filing rules, and admission rosters—form the basis of basis retrieval. Enrollment plans come from the provincial admissions authority's release for the current year, with fields including institution code, major name, planned number, batch, and subject category, and updates are concentrated before volunteer application; filing rules come from admissions documents and institution charters, with fields including filing ratio, ranking score, tie handling, and withdrawal conditions, and supplementary explanations may be added as the batch progresses; admission rosters come from on-site printing and signature confirmation during admission, with fields including candidate number, admitted major, admission date, and handler signature. When the materials enter hybrid retrieval, enrollment plans and filing rules run in parallel as full-text passages and vector chunks, while admission rosters run in parallel as scanned-document text and structured entries; after merging, the candidate scope is large, so a reranking model is needed to reorder by business relevance, and then citation trimming and answer generation produce basis entries that can be traced back to the original text.
Configuration and implementation checks
This step is performed by the reranking model in the retrieval service, which outputs the sorted candidates and rerank scores. "Basis entries that can be traced back to the original text" are then completed by citation trimming and answer generation. Hybrid retrieval simultaneously runs the vector and full-text paths and merges the results; after merging, the scope is large and similarity filtering cannot be applied directly, so the reranking model is usually used again to sort and filter by rerank score.
| Object | Check | Verification action |
|---|---|---|
| Enrollment plans | Institution code, major name, planned number, batch, and subject category must be consistent with the current release, and the date must fall within the current year's batch | Check the two-path switches in the search parameters, confirm that searchMode enables both vector and full-text, see whether candidates come from both paths, and record the number of candidates and scores |
| Filing rules | Filing ratio, ranking score, tie handling, and withdrawal conditions must correspond to the current documents; if the clause version is not synchronized, it must not be used as a basis | Compare the order and scores before and after reranking, use usingReRank to confirm that reranking has participated, and check whether the top few results fall in the original text of the rule clauses |
| Admission rosters | Candidate number, admitted major, admission date, and handler signature must be complete; entries missing a signature or receipt must not be cited back on their own | Control the number of candidates by limit, check whether roster entries are consistent with the two materials of plans and rules in name, quantity, and date, and record the number of candidates and scores |
Three common errors
- Symptom: the admitted major name or planned number for the same candidate does not match between two materials. Cause: the versions of the enrollment plan and admission roster are not synchronized, or an old major catalog was used during entry.
- Symptom: the clause date of the filing rules and the date of the admission roster cross batches. Cause: the current document did not replace the old version, and both versions were hit at the same time during retrieval without date filtering.
- Symptom: an admission roster entry lacks a handler signature or receipt. Cause: the scanned document missed pages or the signature page was not archived with the roster, yet the entry was still hit by retrieval.
How to verify the setup
Acceptance of this function: compare the order and scores of the top three results before and after enabling reranking, confirm that usingReRank is enabled, that changes in the number of candidates and scores can be recorded, and that the similarity threshold is for reference only.
Business test: ask a basis question that spans materials and see whether the answer gives a traceable source.
- Check the two-path switches of
searchModein the search parameters, confirm that both vector and full-text are enabled, and see whether candidates come from both paths at the same time. - Compare the order and scores before and after reranking, control the number of candidates by
limit, and check the tokens of the cited text and the number of array entries and record them separately.