What this category's data looks like
This page addresses exam organization and violation handling scenarios. Materials come from three categories: exam administration manuals, examination room record sheets, and violation fact records, with an update cadence that rolls with exam batches and version revisions. The three material categories play different roles when entering the document parsing and chunking feature: exam administration manuals provide the process and terminology baseline, examination room record sheets carry time, examination room, and personnel fields, and violation fact records carry fact description and signature receipt fields. Field characteristics center on names, quantities, dates, signatures, and receipts. The same record often shows mismatched names, quantities, and dates across two materials, versions not synchronized, and missing signatures or receipts. After parsing, materials are first persisted as visible data rows, then enter the three-state index flow, for retrieval and subsequent processing to generate a unified terminology table and comparison draft.
Configuration and implementation checks
This step is executed by the parsing and training task (server-side task), outputting data rows and index states (parsed, indexing, indexed). The "unified terminology table and comparison draft" is then completed by retrieval and subsequent processing.
| Object | Check | Verification action |
|---|---|---|
| Exam administration manual | Terminology baseline and process fields are complete; names, dates, and signature columns can be traced back to the source text | Check whether the data row state is parsed / indexing / indexed, and check whether terminology is consistent across chunks under chunkSize and chunkSplitter |
| Examination room record sheet | Names, quantities, and dates for the same record are consistent between the manual and the record sheet; signatures and receipts are not missing | Check the file record and the data row separately, confirm the processing mode corresponding to trainingType, and sample three terms to check their spelling throughout the document |
| Violation fact record | Fact description, date, and signature receipt can be traced back to the source text; version is synchronized with the manual | Check whether the data rows covered by the rebuild task contain only completed-index data, and compare names, quantities, and dates one by one |
| Data rows and index states | In one of the three states and consistent with the interface display; data pending indexing does not participate in recall | Open the data rows to check state transitions, confirm that the initial training covers data pending indexing and that rebuild targets only completed-index data |
Three common errors
- Symptom: newly uploaded materials cannot be found. Cause: data rows are in parsed or indexing state; data pending indexing is naturally isolated at the retrieval layer and is not recalled. Successful parsing does not equal a retrieval hit.
- Symptom: names, quantities, and dates for the same record do not match across two materials. Cause: versions are not synchronized, the terminology table is not unified, wording differs across chunks after chunking, and signatures or receipts are missing.
- Symptom: empty data rows remain after a training failure. Cause: training interruption leaves orphan data, states are not used to directly edit or delete, and the rebuild scope mistakenly covers data pending indexing.
How to verify the setup
Feature acceptance: upload one material, observe the data row state transitions and the coverage scope of the rebuild task, confirm the states are parsed / indexing / indexed, confirm that rebuild covers only completed-index data, and check the chunking and processing modes corresponding to chunkSize, chunkSplitter, and trainingType.
Business test: sample three terms to check whether their spelling is consistent throughout the document, whether translated names can be traced back to the source text, and whether names, quantities, dates, signatures, and receipts for the same record align across the manual, record sheet, and violation fact record.
- Open the data rows to check that states are consistent with the interface display, check the file record and data row identity separately, and confirm that the unique data identifier is bound correctly.
- Look at the data rows covered by the rebuild task and confirm that only completed-index data is covered; initial training covers data pending indexing and indexing data, and the two are not interchangeable.