What this category's data looks like
The pre-departure training and overseas safety reminders category contains three types of materials: pre-departure training courseware, safety reminder checklists, and accommodation contact forms. Courseware is updated by semester and includes training topic, date, and instructor signature; safety reminder checklists are updated by destination and include risk items, emergency contact information, and receipt signature; accommodation contact forms are updated by batch and include accommodation name, address, contact person, and confirmation date. When these three types of materials enter knowledge base retrieval, courseware and checklists serve as the basis source role, and accommodation contact forms serve as the contact information verification role. The name, quantity, and date of the same record in the courseware and checklist must be aligned item by item, the version is based on the current date, and if the signature or receipt is missing, it is deemed unusable as a basis. After the materials undergo question optimization, semantic or full-text recall, multi-path merging, reranking, similarity filtering, and citation upper limit trimming, candidate chunks and similarity scores are obtained.
Configuration and implementation checks
This step is executed by the knowledge base retrieval service (retrieval node and search test), outputting candidate chunks and similarity scores (after mode, merging, reranking, filtering, and upper limit trimming). The basis items that can be traced back to the original text are then completed by subsequent nodes or manually.
| Object | Check | Verification action |
|---|---|---|
| Pre-departure training courseware | Fields include training topic, date, and instructor signature; the name, quantity, and date of the same record in the courseware and safety reminder checklist must be aligned item by item, the version takes the current date, and if the signature is missing it cannot be used as a basis | In the knowledge base search test, switch searchMode to semantic, full-text, and hybrid once each, record the number of candidate items and scores, and check whether the hit chunks can be traced back to the courseware original text |
| Safety reminder checklist | Fields include risk items, emergency contact information, and receipt signature; when the version is not synchronized, the version with the current date prevails, and if the receipt is missing it cannot be used as a basis | For the same question, test once each with synonymous rewritten question sentences, compare candidates and scores, and turn usingReRank on and off to see changes in the ranking of the top few items |
| Accommodation contact form | Fields include accommodation name, address, contact person, and confirmation date; the name and date of the same record in the courseware and contact form must be aligned, and if the confirmation receipt is missing it cannot be used as a basis | Check the tokens of the returned citation text and the similarity filtering threshold, record the number of array items separately, and after citation upper limit trimming check whether the key chunks are complete |
| Knowledge base retrieval service | One retrieval consists of question optimization, semantic/full-text/hybrid recall, multi-path merging, reranking, similarity filtering, and citation upper limit trimming; the node only outputs candidates and citation content | In the search test, compare the number of candidate items before and after limit trimming, and confirm that what is returned is candidates and citation content, not statistical results |
Three common errors
- Phenomenon: the name, quantity, and date of the same record in the courseware and safety reminder checklist do not match. Cause: the two materials have different update rhythms, and item-by-item field alignment was not done.
- Phenomenon: the safety reminder checklist version is not synchronized, and old-version risk items are still recalled. Cause: the current version date is not marked, and retrieval does not filter by version.
- Phenomenon: the accommodation contact form lacks a signature or receipt but is still treated as a basis item. Cause: when the materials were ingested, the completeness of the signature and receipt fields was not validated.
How to verify the setup
Acceptance of this feature: test once each with synonymous rewritten question sentences, compare candidates and scores, then turn reranking on and off to see the ranking of the top few items; switch searchMode and adjust limit, record the number of candidate items, scores, and citation text tokens, and record the number of array items separately.
Business test: ask a basis question spanning materials, see whether the answer gives traceable sources, and check whether the sources match the current version and signature/receipt.
- Check the name, quantity, and date fields item by item for the three types of materials, and mark the current version date and signature/receipt status.
- In the knowledge base search test, test once each with synonymous question sentences, record the number of candidate items and scores, and turn reranking on and off to compare the ranking of the top few items.