What this category's data looks like
Enterprise independent skill talent evaluation involves three core types of materials: evaluation filing forms, evaluation specifications, and candidate rosters. The evaluation filing form contains the filing number, occupation name, level, evaluation method, and date signature; the evaluation specification contains assessment modules, weights, scoring rules, and version number; the candidate roster contains name, ID number, trade, batch, and employer seal. Materials are collected by enterprise evaluation agencies by batch, and the update cadence is triggered by filing changes, specification revisions, and roster supplements. When entering knowledge base retrieval, the filing form is used to lock the current version and filing status, the evaluation specification is used to recall assessment requirements and scoring clauses, and the roster is used to verify personnel fields and batch dates. The fields of the three material types intersect, and names, quantities, dates, and signature receipts need to be compared item by item in the retrieved candidate chunks to ensure the current version is unique, old versions are traceable, and citations are synchronized.
Configuration and implementation checks
The knowledge base retrieval service (retrieval node and search test) only outputs candidate chunks and similarity scores (after mode, merge, rerank, filter, and limit trimming); the version ledger and change records must be handled by subsequent nodes or manually.
| Object | Check | Verification action |
|---|---|---|
| Evaluation filing form | Whether the filing number, occupation name, level, evaluation method, and date signature are consistent with the current version | In the search test, use synonymous questions to retrieve the filing number, check the candidate chunks and scores, and inspect the result differences after switching searchMode |
| Evaluation specification | Whether the assessment modules, weights, scoring rules, and version number are unique and old versions are traceable | Enable usingReRank to compare the ranking of the top few results, confirm that current-version chunks rank at the front, and that old-version chunks can be recalled separately |
| Candidate roster | Whether name, ID number, trade, batch, and employer seal match the filing form and specification | Use the batch number to retrieve roster candidates, record the number of candidates and scores, and check whether key fields are retained after similarity filtering |
| Version ledger and change records | Current version is unique, old versions are traceable, citations are synchronized, and signatures and receipts are complete | After changing one piece of content, retrieve again to see whether the version number and citations are updated synchronously, and record the number of cited items and text tokens after limit trimming |
Three common errors
- Symptom: The name and quantity of the same record do not match between the filing form and the roster. Cause: The two materials are maintained separately, and field-level comparison has not been performed against the current version as the baseline.
- Symptom: The evaluation specification has been revised, but retrieval still recalls old-version clauses. Cause: The version is not synchronized, old-version chunks are not marked invalid, and candidates are not filtered by version.
- Symptom: The filing form or roster lacks signatures or lacks receipts. Cause: When materials were collected, the signature fields and receipt attachments were not checked, and retrieval candidates only present body chunks.
How to verify the setup
This feature acceptance: test once each with synonymously rewritten questions, compare candidate chunks and similarity scores, then toggle rerank on and off to inspect the ranking of the top few results, and confirm the changes in the number of candidates and scores under searchMode and usingReRank.
Business test: change one piece of content, see whether the version number and citations are updated synchronously, and check and record separately the cited text tokens and array item count after similarity filtering.
- In the search test, switch
searchModeand retrieve the same filing number using semantic, full-text, and hybrid modes respectively, comparing the number of candidates and scores. - Compare the ranking of the top few results before and after enabling
usingReRank, confirming that current-version chunks stably rank at the front and old-version chunks can be recalled separately.