What this category's data looks like
Technical college academic education management materials mainly consist of student registration forms, teaching plans, and skills training records, all three of which are updated on a rolling basis according to semester and academic year milestones. Student registration forms carry student identity, registration date, and approval signature; teaching plans carry course names, credit hours, and execution dates; skills training records carry training items, completion counts, and instructor receipts. When these three types of materials enter the knowledge base retrieval and recall function, they each participate as independent recall sources in candidate generation, and field differences and version differences will be exposed simultaneously in the candidate snippets. Therefore, it is necessary to compare item by item according to material name, field name, date, and signature, to avoid the same record having mismatched names, quantities, and dates across different materials.
Configuration and implementation checks
This step 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 upper-limit trimming). The basis entries that can be traced back to the original text are then completed by subsequent nodes or manually.
| Object | Check | Verification action |
|---|---|---|
| Student registration form | Fields include student name, registration date, and approval signature; the name, quantity, and date of the same record in two materials must be aligned item by item | In the search test, set searchMode to hybrid, record the number of candidates and scores, and check whether the registration date and signature fall in the same snippet |
| Teaching plan | Fields include course name, credit hours, and execution date; when versions are not synchronized, the same course name corresponds to different credit hours | Switch searchMode to semantic and full-text and test each once, compare the number of candidates and scores, and check whether the course name and credit hours are consistent |
| Skills training record | Fields include training item, completion count, and instructor receipt; if the signature or receipt is missing, the entry cannot serve as a basis | Compare the ranking of the top few entries before and after enabling usingReRank, filter low-scoring candidates by similarity, and check the receipt and completion count |
| Retrieval result citation | The number of citations is the array length, and the citation content budget is checked separately by tokens; the two are not compared for equality | Record the number of candidates and scores, record the citation text tokens separately, and check whether the key snippets are retained |
Three common errors
- Symptom: The same registration record has mismatched names and dates in the student registration form and the teaching plan. Cause: The two materials are entered separately and versions are not synchronized, so retrieval hits snippets from different versions.
- Symptom: The completion count in the skills training record is inconsistent across two materials. Cause: The training record and the teaching plan were not updated according to the same date milestone, and the candidate snippets come from an old version.
- Symptom: A teaching plan is still recalled even though the signature or receipt is missing. Cause: Similarity filtering only trims by text relevance and does not verify the signature field, so manual tracing back to the original text is required.
How to verify the setup
Acceptance of this function: test once each with synonym-rewritten questions, compare the number of candidates and scores, then toggle rerank to observe the ranking of the top few entries, and check the actual values of searchMode, similarity, and usingReRank.
Business test: ask a cross-material basis question and see whether the answer provides a traceable source.
- In the knowledge base search test, switch
searchModeamong semantic, full-text, and hybrid, record the number of candidates and scores, and check whether the candidates include both semantic and full-text results. - Run once with
usingReRankenabled and once disabled, record the ranking changes of the top few entries bysimilarity, and separately record the citation text tokens and the array length.