What this category's data looks like
In the graduate admissions re-examination and admission scenario, basis queries must deal with three types of materials: the re-examination plan, the candidate score table, and the transfer list. The re-examination plan provides the process, weights, and time criteria; the candidate score table carries the preliminary and re-examination scores, rankings, and signature fields; the transfer list records the transfer batch, proposed admission status, and receipt. When the materials enter the vector model and indexing stage, the re-examination plan is chunked as rule-based text to generate a text index; the candidate score table and transfer list enter as table rows or records, with fields including name, ID, date, quantity, and signature. The update rhythm rolls forward with batch progress and list publication; the name, quantity, and date of the same record in the two materials may not match, and when versions are not synchronized or signatures or receipts are missing, the index may still recall, but whether a basis item can point back to the original text depends on subsequent nodes.
Configuration and implementation checks
This stage is executed by the indexing service and the vector database, outputting vectors and index status (whether recall is possible, whether data in the same group is merged). "Basis items that can point back to the original text" are completed next by subsequent AI nodes, code nodes, or manual work (retrieval and citation trimming only take supporting materials).
| Object | Check | Verification action |
|---|---|---|
| Re-examination plan | Fields include process nodes, weights, dates, and signature fields, and the version must be current; when the name and date of the same record in the two materials do not match, it cannot be used directly as a basis | In the knowledge base settings, check that the index model is consistent with the vector model actually used by the knowledge base, open a data row to check whether the index fields cover all vector identifiers, and run a retrieval to see whether the same group is merged |
| Candidate score table | Fields include preliminary score, re-examination score, ranking, date, and signature; quantities and names must be aligned item by item with the transfer list, and items missing signatures or receipts are marked separately | Check the data scope covered by the rebuild task, and confirm that after changing models, only data that has already been indexed is triggered; for items not recalled, check the data row index fields and the defaultToken, maxToken budgets |
| Transfer list | Fields include transfer batch, proposed admission status, date, and receipt; when the quantity and date of the same record in the two materials do not match, point back according to the current version | Open a data row to check the index fields, after retrieval merge back to the original data by vector identifier, and take the highest score for the same group of data; record the number of citation items and citation text tokens, and record the number of array items separately |
| Data row index fields | Multiple vector identifiers of one piece of data must all be present, and the status must be indexed or findable in the rebuild task record | Open a data row to check the index fields and confirm that multiple vector identifiers are complete; after changing models, check the rebuild task record to confirm the coverage scope |
Three common errors
- Symptom: duplicate items appear in retrieval results. Inspect the following: the same group of original data corresponds to multiple vector identifiers, and merging was not done by highest score, causing the name and date of the same record in the two materials to not match.
- Symptom: recall quality drops after changing models. Inspect the following: the rebuild task is triggered only for data that has already been indexed, materials whose versions are not synchronized are still recalled by the old index, and items missing signatures or receipts are mixed in.
- Symptom: a data row has content but cannot be retrieved. Inspect the following: the index fields do not record all vector identifiers, or the index model is inconsistent with the vector model actually used by the knowledge base, and when dates and quantities do not match, it cannot point back.
How to verify the setup
This feature's acceptance: sample a data row and look at the index fields, confirm that multiple vector identifiers are all present and the status is indexed or findable in the rebuild task record; then run a retrieval to see whether the same group is merged, and take the highest score for the same group of data.
Business test: ask a basis question across materials and see whether the answer gives a traceable source; check the number of citation items and citation text tokens, record the number of array items separately, and quoteMaxToken is only used as the citation content budget.
- In the knowledge base settings, check that the index model is consistent with the vector model actually used by the knowledge base, and check the data scope covered by the rebuild task.
- Open a data row to check the index fields, confirm that multiple vector identifiers are complete, after retrieval merge back to the original data by identifier, and take the highest score for the same group of data.