What this category's data looks like
This page addresses the category of pre-departure training and overseas safety reminders. The core materials include pre-departure training courseware, safety reminder checklists, and accommodation contact lists. These materials are usually updated by batch, and their fields cover name, date, quantity, signature, and receipt. When entering the knowledge base retrieval and recall function, the training courseware plays the role of the basis for policies and procedures, the safety reminder checklist plays the role of risk items and emergency guidance, and the accommodation contact list plays the role of specific addresses and contact information. In retrieval, the three types of materials each serve as sources of candidate chunks, which are recalled by the retrieval service through semantic and full-text channels, and then merged, reranked, filtered, and truncated by limit to form cited content, for subsequent nodes or humans to complete the discrepancy list.
Configuration and implementation checks
This step is performed by the knowledge base retrieval service (retrieval node and search test), which outputs candidate chunks and similarity scores (after mode, merge, rerank, filter, and limit truncation). The discrepancy list is then completed by subsequent nodes or humans.
| Object | Check | Verification action |
|---|---|---|
| Pre-departure training courseware | The name, date, and signature fields must be consistent with the safety reminder checklist | In the search test, test once each with synonymous questions, record the number of candidates and scores, and switch searchMode to compare semantic, full-text, and hybrid modes |
| Safety reminder checklist | The quantity, date, and receipt fields must be consistent with the accommodation contact list | Turn reranking on and off, compare the order of the top few candidates, and check the impact of usingReRank on ranking |
| Accommodation contact list | The name, quantity, and date of the same record match across the two materials | Observe whether the candidates include both semantic and full-text results, and record the chunks retained after similarity filtering |
| Knowledge base retrieval service | Outputs candidate chunks and similarity scores, not statistical tables or business documents | Check the number and content completeness of returned citations, record the tokens of the cited text separately, record the number of array items separately, and note that limit only controls the citation upper limit |
Three common errors
- Symptom: The name, quantity, and date of the same record do not match between the training courseware and the safety reminder checklist. Cause: The two materials have different update batches, and no field-level comparison was performed.
- Symptom: The accommodation contact list and the ledger version are not synchronized, and old addresses are still recalled. Cause: The knowledge base is not marked by version, and retrieval hits old chunks.
- Symptom: The safety reminder checklist lacks signatures or receipts, and the discrepancy list cannot be closed. Cause: The signature column was not checked during material archiving, and receipts were not attached.
How to verify the setup
This function's acceptance: test once each with synonymously rewritten questions, compare candidates and scores, then turn reranking on and off to inspect the order of the top few; record the number of candidates and scores, and check the effects of searchMode, usingReRank, similarity, and limit.
Business test: use one real record to find fields inconsistent with the ledger, and check name, quantity, date, signature, and receipt.
- In the search test, switch
searchModeto compare the number of candidates and scores across semantic, full-text, and hybrid modes. - Turn reranking on and off, compare the order of the top few candidates, and record the chunks retained after
similarityfiltering.