What this category's data looks like
This category's data covers three types of materials: employment agreements, graduation destination registration forms, and employer certificates. An employment agreement contains the employer name, Unified Social Credit Code, position, signing date, and signatures and seals of both parties; a graduation destination registration form contains the destination type, employer location, registration certificate information, and personal confirmation; an employer certificate contains the offered position, onboarding date, employer contact, and receipt. The three types of materials come from different sources and update at different rhythms, with concentrated modifications, supplementary entries, and re-signings frequent during signing season. After document parsing and chunking, the employment agreement serves as the primary basis for signing, the graduation destination registration form serves as the source for statistical criteria, and the employer certificate serves as supplementary evidence; each becomes an independent data row bound to a unique data identifier for subsequent retrieval and version comparison.
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 version ledger and change records are then completed by retrieval and subsequent processing. After parsing is complete, parsed chunks 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; and when the model is changed, only data with completed indexing is rebuilt.
| Object | Check | Verification action |
|---|---|---|
| Employment agreement | Fields contain employer name, position, signing date, and signatures and seals of both parties; the date matches the registration form | Open the data row to check whether the status is parsed, indexing, or indexed, and compare signatures and dates |
| Graduation destination registration form | Destination type, employer location, registration certificate information, and personal confirmation are complete; the name matches the agreement | Check the file record and the data row separately, confirm that identity identifiers are distinguishable, and inspect the chunking results of chunkSize and chunkSplitter |
| Employer certificate | Offered position, onboarding date, employer contact, and receipt are complete; the quantity matches the signing records | Check the receipt and signature, and view the processing mode corresponding to trainingType and the index status |
| Version ledger and change records | The current version is unique, old versions are traceable, references are synchronized; after changing one piece of content, the version number and references are updated in sync | Check whether the data rows covered by the rebuild task contain only data with completed indexing |
Three common errors
- Symptom: newly uploaded materials cannot be found. Cause: the data row is still in the pending-indexing state and does not participate in recall before indexing is complete; successful parsing does not equal a retrieval hit.
- Symptom: the same record's employer name, number of positions, or signing date does not match between the agreement and the registration form. Cause: the two materials are parsed and persisted separately, lacking cross-material field comparison and version synchronization.
- Symptom: empty data rows remain after a training failure, or rebuild time is abnormal after changing the model. Cause: failed data is not edited or deleted in time, and the rebuild scope mistakenly covers data pending indexing.
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; the status should go from parsed to indexing to indexed, and the rebuild should cover only data with completed indexing.
Business test: change one piece of content and see whether the version number and references are updated in sync; the number of references is recorded by array length, the tokens of reference text are checked separately, and quoteMaxToken is the reference content budget; the two are not compared for equality.
- Open the data row to check whether the status matches what the interface displays, confirm one of the three states, and check the file record and data row identifiers separately.
- View the data rows covered by the rebuild task and confirm they contain only data with completed indexing; directly edit or delete failed training data to avoid residual empty data rows.