What this category's data looks like
In the filing and implementation stages, enterprise independent skilled talent evaluation generates three types of materials: the enterprise evaluation filing form, the evaluation specification, and the candidate roster. The filing form carries the evaluation agency, occupational direction, and filing date; the evaluation specification states the assessment modules, skill requirements, and scoring rules; the roster records candidate identity, declared occupation, and batch. The three types of materials come from different sources and have inconsistent update rhythms: the filing form is adjusted with filing changes, the evaluation specification is reversioned with standard revisions, and the roster is added to or supplemented by batch. Field characteristics center on names, quantities, dates, and signature receipts: the same candidate's name spelling, declared number of people, and batch date in the roster and the filing form need to align, and the version numbers and seals of the evaluation specification and the filing form need to correspond. When the materials enter the document parsing and chunking stage, each is split into data rows and bound to a unique data identifier, becoming the input for subsequent retrieval and discrepancy localization.
Configuration and implementation checks
This step is executed by the parsing and training task (server-side task), outputting data rows and index status (parsed, indexing, indexed). The "discrepancy list" is completed next by retrieval and subsequent processing. After parsing is complete, parsing and chunking are immediately persisted as visible data rows and bound to a unique data identifier; the index transitions through the three states parsed, indexing, and indexed; data pending indexing does not participate in recall; changing the model only rebuilds data that has completed indexing.
| Object | Check | Verification action |
|---|---|---|
| Enterprise evaluation filing form | The filing agency name, evaluated occupation, filing date, and seal must be consistent with the evaluation specification | Check the agency name, occupation name, filing date, signature, and receipt in the filing form and the evaluation specification, confirm the version numbers are synchronized; when chunking, use chunkSize and chunkSplitter to fix field boundaries |
| Evaluation specification | The assessment module names, skill requirements, scoring rules, and version number must be consistent with the filing form | Check the module names, quantities, effective dates in the specification against the corresponding fields in the filing form, and check whether the old version remains after reversioning; use trainingType to distinguish vector and QA processing modes |
| Candidate roster | Candidate names, declared occupation, number of people, and batch date must be consistent with the filing form | Compare the name spelling, declared number of people, and batch date in the roster and the filing form one by one, check entries missing signatures or receipts, and record inconsistent fields |
| Data rows and index status | Is in one of parsed, indexing, indexed and consistent with the interface display | Open the data rows to check the status, separately check the identity identifiers of the file record and the data row, and check whether the data rows covered by the rebuild task contain only data that has completed indexing |
Three common errors
- Symptom: newly uploaded materials cannot be found. Cause: the data rows are in parsed or indexing status; data pending indexing is naturally isolated at the retrieval layer, and successful parsing does not mean retrieval will necessarily hit.
- Symptom: the same candidate's name, number of people, and batch date do not match across two materials. Cause: the filing form and roster versions are not synchronized, old-version fields were not updated with reversioning, and discrepancies were not exposed before review.
- Symptom: empty data rows remain after training failure, or rebuild takes abnormally long after changing the model. Cause: failed data was not edited or deleted in time; the rebuild scope mistakenly covers data pending indexing, and should only target data that has completed indexing.
How to verify the setup
Acceptance of this feature: upload one material, observe the data row status transition from parsed to indexing to indexed, confirm that data pending indexing does not participate in recall, and check whether the data rows covered by the rebuild task contain only data that has completed indexing.
Business test: use one real record to find fields inconsistent with the ledger, check names, quantities, dates, and signature receipts, form a discrepancy list, submit it to the reviewer for confirmation, and write it back.
- Open the data rows to check whether the status is consistent with the interface display, separately check the identity identifiers of the file record and the data row, and confirm binding to a unique data identifier.
- Check whether the rebuild task coverage contains only data that has completed indexing, and record the tokens of the referenced text and the number of array entries separately, avoiding mixing the number of references with the reference content budget.