What this category's data looks like
Vocational college skills competition organization involves three types of materials: competition regulations, registration forms, and training plans. Regulations are issued by the organizer and updated annually with the competition event; registration forms are submitted by participating schools and contain contestant information, group category, and signature; training plans are written by instructors and contain phase goals, dates, and coach signature. When these three types of materials enter hybrid retrieval, regulations serve as the rule basis, registration forms serve as personnel and qualification evidence, and training plans serve as process records, each bearing a different reference-back role. At the field level, attention is paid to whether names, quantities, dates, signatures, and receipts are consistent. In terms of update cadence, regulations change versions by competition year, registration forms are submitted by registration batch, and training plans roll forward by training cycle. When the same record's name, quantity, or date does not match across two materials, the version is not synchronized, or a signature or receipt is missing, these are the main risk points for whether a basis entry can reference 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 rerank scores. The basis entries that can reference 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 then used to sort and filter by rerank score.
| Object | Check | Verification action |
|---|---|---|
| Competition regulations | Names, dates, and seals must be consistent with the current version, and entries must be able to reference back to the original text | Check the regulation version number and release date, spot-check the position of entries in the original text, and confirm whether both paths' candidates hit under searchMode |
| Registration form | Contestant names, quantities, group category, and signature receipt must match the regulations and training plan | Compare the group category names and headcounts in the registration form and regulations, check whether signatures and receipts are complete, and record the number of candidates and rerank scores |
| Training plan | Phase dates, training volume, and coach signature must be consistent with the registration form and regulations | Check the training plan dates against the registration form batch dates, check the coach signature, and confirm that usingReRank has been used for sorting and filtering |
| Rerank results | The top few candidates are relevant to the basis question, and the scores can be used for filtering | Compare the order and scores of the top three before and after reranking, adjust limit to observe changes in the number of candidates, and confirm that similarity is not used as a direct filtering basis |
Three common errors
- Symptom: there are many candidates but the top few are irrelevant. Cause: after hybrid retrieval merging, the scope is large, rerank scores are not used for filtering, and similarity filtering appears to be ineffective.
- Symptom: the same contestant's name, quantity, or date does not match across two materials. Cause: the registration form and training plan are each updated separately, versions are not synchronized, and signatures or receipts are missing.
- Symptom: results are clearly biased toward keyword hits. Cause: the full-text path has too high a weight, candidates from the vector path are drowned out, and rerank scores do not participate in final trimming.
How to verify the setup
Acceptance of this feature: compare the order and rerank scores of the top three before and after enabling reranking, confirm changes in the number of candidates and scores, and check whether usingReRank participates in sorting and filtering.
Business test: ask a basis question that spans materials and see whether the answer gives a referenceable source.
- In the search parameters, check the switches for the vector and full-text paths, and confirm that both paths are enabled and merged under
searchMode. - Compare the sorting and scores before and after reranking, and check whether the values of
limitandsimilaritymake the number of candidates and scores explainable.