What this category's data looks like
This page addresses the enterprise new employee onboarding training scenario, conducting business basis queries around three types of materials: onboarding training plans, course sign-in sheets, and job descriptions. Onboarding training plans typically contain the course name, training date, responsible department, and approval signature, with a finalization cadence before each session begins; course sign-in sheets contain attendees, sign-in date, course name, and personal signature, and are generated during class; job descriptions contain the job title, responsibility items, qualification requirements, and version number, and are synchronized with organizational adjustments. When these three types of materials enter the knowledge base retrieval and recall function, they respectively serve as the basis for course arrangements, the basis for attendance records, and the basis for scope of responsibilities. During retrieval, the recall channel is selected according to the material form, and candidate snippets and similarity scores are handed to subsequent nodes or humans to form basis entries that can be traced back to the original text.
Configuration and implementation checks
This step is performed by the knowledge base retrieval service (retrieval node and search test), outputting candidate snippets and similarity scores (after mode, merge, rerank, filter, and upper-limit trimming). Basis entries that can be traced back to the original text are then completed by subsequent nodes or humans.
| Object | Check | Verification action |
|---|---|---|
| Onboarding training plan | Whether the course name, training date, responsible department, and approval signature are complete, and whether they match the sign-in sheet | In the search test, switch searchMode and check whether the candidate snippets cover both the plan and the sign-in sheet, and record the number of candidates and scores |
| Course sign-in sheet | Whether the attendees, sign-in date, course name, and personal signature are complete, and whether they are consistent with the plan date | Test once each with synonymous questions, compare candidates and scores, and check whether the returned snippets can be traced back to the original text of the sign-in sheet |
| Job description | Whether the job title, responsibility items, qualification requirements, and version number are the current version, and whether they are consistent with the job title in the plan | Turn usingReRank on and off to compare the ranking of the first few items, and check whether the version number falls within the returned snippets |
| Cited content | Whether key snippets of the original text are still retained after similarity filtering and citation upper-limit trimming | Record the number of array items in the returned citations, check the tokens of the citation text separately, and do not compare the number of items with the budget as equal |
Three common errors
- Symptom: The name and date of the same course do not match across the two materials. Cause: The materials are maintained separately, were not checked against the current version at entry, and are cited directly after retrieval hits.
- Symptom: The job description version is not synchronized, and old responsibility items are still recalled. Cause: The update cadence is inconsistent with the session start cadence, and the knowledge base does not filter by version.
- Symptom: The sign-in sheet lacks a signature or receipt, but the answer still gives that basis. Cause: Retrieval only outputs candidates and scores, the signature field is not checked, and humans do not trace back to the original text.
How to verify the setup
Acceptance of this feature: test once each with synonymously rewritten questions, compare candidates and scores, then turn usingReRank on and off to see the ranking of the first few items; at the same time, switch searchMode to observe differences in candidate sources, and record the number of candidates and scores.
Business test: ask a basis question that spans materials, and see whether the answer gives a traceable source.
- In the search test, switch
searchMode, compare the number of candidates and scores under the three modes of semantic, full-text, and hybrid, and confirm that the candidate snippets can be traced back to the original text. - Run once each with
usingReRankon and off, compare the ranking of the first few items; record the number of array items in the returned citations, and check the tokens of the citation text separately.