What this category's data looks like
University employment signing and graduation destination statistics involve three types of materials: employment agreements, graduation destination registration forms, and employer certificates. An employment agreement contains the employer name, unified social credit code, position, signing date, and signatures of both parties; a graduation destination registration form contains the destination category, employer name, reporting date, and personal confirmation; an employer certificate contains the acceptance opinion, contact person, date of issuance, and seal. The materials come from departmental summaries, university-level reviews, and employer receipts, are updated by batch, and follow a dense rhythm during signing season and dispatch period. When the three types of materials enter knowledge base retrieval, each is recalled as an independent fragment: agreement fragments correspond to signing fields, registration form fragments correspond to destination fields, and certificate fragments correspond to acceptance fields, for subsequent nodes to compare whether the name, quantity, and date of the same record are consistent.
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 pattern, merge, rerank, filter, and limit trimming). The issuable material drafts and handover records are completed by subsequent nodes or manually.
| Object | Check | Verification action |
|---|---|---|
| Employment agreement | The employer name, position, signing date, and signatures of both parties fields are complete, and the name and date are consistent with the same record in the registration form | In the search test, search once each using different phrasings of the same record, check the number of candidates and the scores, and switch searchMode to compare the differences between semantic and full-text candidates |
| Graduation destination registration form | The destination category, employer name, reporting date, and personal confirmation are complete, and the version and issuance date are synchronized | Compare the order of the top few items before and after enabling usingReRank, and confirm that the key fragments of the registration form are not trimmed out by similarity filtering |
| Employer certificate | The acceptance opinion, contact person, date of issuance, and seal are complete, and the receipt is traceable | Adjust limit to observe changes in the number of candidates, record the array length separately, and then check whether the tokens of the referenced text retain the date of issuance and seal description |
| Handover record | The responsible person, receipt time, version number, and material list are consistent, and missing items can be blocked | Draft one according to the template and complete one handover, checking whether it is blocked when a signature or receipt is missing |
Three common errors
- Symptom: The employer name, quantity, and date of the same record do not match between the agreement and the registration form. Cause: The two fragments were recalled separately without field-level comparison and went directly into drafting.
- Symptom: The material versions are not synchronized, and an old registration form is mixed with a newly signed agreement. Cause: Retrieval was not filtered by issuance date, and candidate fragments were mixed across versions.
- Symptom: The handover record lacks a signature or receipt, but the next step still signs off. Cause: The missing-item validation was not bound to the responsible person field, and the candidate fragments did not cover the sign-off information.
How to verify the setup
This feature's acceptance: search once each using synonymously rewritten questions and compare the number of candidates and scores; then toggle rerank on and off to see whether the order of the top few items changes, and check the actual values of similarity, limit, usingReRank, and searchMode.
Business test: draft one according to the template and complete one handover, and see whether missing items are blocked.
- In the knowledge base search test, switch
searchMode, search once each for different phrasings of the same record, and record the number of candidates and scores. - Run once each with
usingReRankenabled and disabled, compare the order of the top few items, and check whether the tokens of the referenced text undersimilarityandlimitretain the key fields.