What this category's data looks like
The written and interview registration forms, Putonghua certificates, and medical examination forms involved in teacher qualification exams and certification are the category materials most often referenced back when replying to parents or writing them into proposals. Written and interview registration forms record the exam stage, subject, exam district, and payment status, with fields including name, ID number, date, and signature; Putonghua certificates record the grade, test time, and certificate number; medical examination forms record the medical examination conclusion, hospital opinion, and physician signature. The update cadence of the materials varies: registration forms are generated by batch, Putonghua certificates are issued by test batch, and medical examination forms have a validity period. When the three enter hybrid retrieval, the registration form, Putonghua certificate, and medical examination form each serve as independent candidate sources, participating in vector and full-text dual-path recall and merging, then after reranking are filtered by score, so that citation trimming and answer generation can produce basis items that can be traced back to the original text.
Configuration and implementation checks
This step is executed by the reranking model in the retrieval service, which outputs the sorted candidates and reranking scores. The basis items that can be traced back to the original text are then completed by citation trimming and answer generation. Hybrid retrieval simultaneously runs vector and full-text dual paths and merges the results; after merging, the scope is large and similarity filtering cannot be done directly, so a reranking model is usually then used to sort and filter by reranking score.
| Object | Check | Verification action |
|---|---|---|
| Written and interview registration form | Fields, date, and signature are complete, and it can be matched to the Putonghua certificate and medical examination form for the same person and same batch | Check the searchMode dual-path switches in the search parameters to confirm that both vector and full-text paths are enabled; compare the ordering and scores before and after reranking to see whether candidates come from both paths |
| Putonghua certificate | The grade, test time, and certificate number are consistent with the registration form and medical examination form records, and the version is current | Check whether usingReRank is bound and used for sorting and filtering; inspect the number of candidates and the scores to confirm that the top few are relevant to the basis question |
| Medical examination form | The medical examination conclusion, hospital opinion, and physician signature are complete, the date has not crossed the validity period, and the receipt is not missing | Check whether the similarity filter takes effect according to the reranking score; check the number of candidates controlled by limit to confirm that the top few can be matched to the original text |
| Hybrid retrieval candidates | The two paths' results have been merged, and candidates come from both the vector and full-text paths | Compare the ordering and scores before and after reranking to confirm that after reranking the top few can still be traced back to the original text |
Three common errors
- Symptom: The name, quantity, and date of the same record do not match between the registration form and the Putonghua certificate. Cause: the two recall paths each hit different fields, and after merging they were not filtered by reranking score, so the number of candidates is large and the top few are irrelevant.
- Symptom: The medical examination form version is not synchronized, and old conclusions are mixed with current requirements. Cause: after hybrid retrieval merging, the scope is large and similarity filtering cannot be done directly; if reranking score filtering is not used, old versions will still rank at the front.
- Symptom: The registration form lacks a signature or the medical examination form lacks a receipt, yet the basis item is still cited. Cause: citation trimming and answer generation only take candidates by reranking score and do not trace back to the original text to check the signature and receipt fields.
How to verify the setup
Acceptance for this feature: compare the order and scores of the top three before and after enabling reranking, and confirm the changes in the number of candidates and scores; check whether the searchMode dual-path switches, usingReRank binding, and similarity filtering take effect according to the reranking score, and that limit only controls the number of candidates.
Business test: ask a basis question that spans materials and see whether the answer gives a traceable source.
- Check the vector and full-text dual-path switches in the search parameters, confirm that both paths of hybrid retrieval are enabled, and see whether candidates come from both paths.
- Compare the ordering and scores before and after reranking, check that the reranking model is bound and used for sorting and filtering, and confirm that the top few can be traced back to the original text.