What this category's data looks like
This page covers three types of materials in the enterprise sales product training scenario: product manuals, sales sales terminology, and channel training schedules. The sources are periodic updates from the sales operations and channel management departments, and the update cadence is mainly quarterly version revisions and ad hoc patches. The field characteristics center on product name, specification quantity, effective date, version number, and signature receipt column. When product manuals enter parsing and chunking, they serve as the primary data source and carry the baseline for object names and parameters; sales sales terminology serves as wording material, used to compare product designations, quantity commitments, and date expressions in the sales terminology; channel training schedules serve as scheduling material, used to verify training sessions, covered products, and signature receipts. After parsing and chunking, the three types of materials each land as visible data rows and are bound to a unique data identifier, for use in 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 then completed by retrieval and subsequent processing. After parsing is complete, parsing and chunking immediately land in the database as visible data rows and are 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, and changing models only rebuilds data whose indexing has already completed.
| Object | Check | Verification action |
|---|---|---|
| Product manual | Name, quantity, date, and version number are consistent within the manual and with the ledger, and the signature receipt column has no missing items | Check the file record and the data row separately, open the data row to check the status, and confirm that the chunks corresponding to chunkSize and chunkSplitter have not truncated fields |
| Sales sales terminology | The product designations, quantity commitments, and effective dates in the sales terminology align item by item with the product manual, the version number is in sync, and there are no missing signatures or receipts | Use the product manual as the baseline for field-by-field comparison, and check whether the processing mode marked by trainingType covers the multi-value expressions in the sales terminology |
| Channel training schedule | The schedule sessions, covered products, dates, and signature receipts are consistent with the ledger, and items whose versions are not in sync are marked separately | Look at the data rows covered by the rebuild task, and confirm that the schedule data rows are in indexed before including them in the comparison scope |
| Ledger record | The three fields of name, quantity, and date for the same object are consistent between the material and the ledger | Use one real record to find the field that is inconsistent with the ledger, write it back to the discrepancy list, and hand it to the reviewer |
Three common errors
- Symptom: a material that was just uploaded cannot be found. Cause: the data row is in parsed or indexing status, data pending indexing does not participate in recall, and it must wait until it transitions to indexed.
- Symptom: an empty data row is left after training fails. Cause: successful parsing does not mean retrieval will definitely hit, failed data is not cleaned up, and it must be directly edited or deleted.
- Symptom: rebuild takes abnormally long after changing models. Cause: the rebuild scope mistakenly covers data pending indexing, when in fact it should only be triggered for data whose indexing has already completed.
How to verify the setup
This feature's acceptance: upload one material, observe the data row status transitions and the coverage scope of the rebuild task, and confirm that the statuses are parsed, indexing, and indexed in order, and that the rebuild only covers indexed data rows.
Business test: use one real record to find the field that is inconsistent with the ledger, record the discrepancy items in name, quantity, or date, and annotate the version and signature situation.
- Open the data row to check the status, confirm it is consistent with the interface display, check the file record and the data row separately, and inspect the chunks and processing modes corresponding to
chunkSize,chunkSplitter, andtrainingType. - Look at the data rows covered by the rebuild task, confirm that only indexed data is covered, organize the inconsistent items in name, quantity, and date as well as the missing signature receipt items into a discrepancy list, hand it to the reviewer for confirmation, and write it back.