What the documents look like
Producer export tax treatment combines the offset of domestic-sale tax amounts with the export refund calculation. Operating reviews and anomaly retrospectives therefore have to reconcile domestic sales, export sales and input tax data at the same time, which is a wider comparison than an export trading company faces.
The main data sources are the electronic tax bureau, ERP sales records, customs electronic information and the financial ledger. The documents handled day to day are the exemption, offset and refund declaration detail schedule, export contracts, the sales detail ledger and export customs declarations. Data updates monthly, so a review must lock a single period and align the declaration detail schedule, the sales detail ledger and the customs declarations item by item. Any cross-period or cross-entity stitching distorts where an anomaly is located.
In the retrieval step this matters because the materials that carry the answer are structured schedules and ledgers rather than narrative policy text. A question about a mismatch between the refund amount in the declaration detail schedule and export sales in the sales detail ledger has to be matched against the same entity, the same period and the same field before any conclusion is drawn.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases with different vector models cannot be associated together, so the vector model identifier is checked first and non-matching bases are excluded. Search modes are compared against test questions across semantic search, full-text search and hybrid search; no single mode suits all materials.
The reference limit is measured in tokens. The minimum relevance value applies only under semantic search or when result reranking is used. Question optimization performs coreference resolution and question expansion, so the original question and the optimized question are compared to check whether the subject or period was altered. Retrieval and similarity scores do not perform permission or entity isolation, and recalled content does not constitute authorization.
Hypothetical paper exercise: a tax accountant at a producer enterprise is reviewing the July 2025 period and intends to enter the question "the exemption, offset and refund amount in the declaration detail schedule does not match export sales in the sales detail ledger, what is the cause". The draft sentence to be judged is: "A retrieval hit on a declaration detail schedule clause means the exemption, offset and refund amount for that month is calculated correctly and a correction filing can be submitted directly." The subject is a producer enterprise, the objects are the declaration detail schedule and the sales detail ledger for the July 2025 period, the status is pending verification, and the authorization boundary is internal review only, with no submission or correction filing.
Configuration and implementation checks
| Configuration or check item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate one exemption, offset and refund base with a consistent vector model identifier | Check the identifier first, exclude bases with different vector models, and do not run cross-model retrieval |
| Search mode | Hybrid search | Run the same question through semantic search and full-text search separately and compare the retrieved content |
| Reference limit | Starting point set to cover the tokens needed for one declaration detail schedule fragment | Test content completeness with short and long materials, then adjust |
| Minimum relevance | Starting point at a middle value | Label relevant and irrelevant samples first, then compare the filtering effect |
| Question optimization | Enabled, comparing the original and optimized question | Check whether the subject and period were rewritten |
| Output reference content | Checked as an array, length may be 0 | The empty-result path still runs; downstream design accounts for an empty array |
Three common errors
- Treating a retrieval hit as evidence that the tax amount calculation is correct. The input is the draft sentence above, the comparison object is the same field in the declaration detail schedule and the sales detail ledger, and the observable result is that the two materials do not show mutually exclusive values for that field; the draft sentence is not supported.
- Forcing knowledge bases with different vector models into one association. The input is the vector model identifier of the two bases, the comparison is whether the identifiers match, and the observable result is that a mismatch leads to exclusion; missed recall is not compared.
- Treating a pending correction request as already executed. The input is the status field of that request, the comparison is between the status and the execution record, and the observable result is that execution evidence is absent and the item is recorded as pending verification.
How to verify the setup
- Enter the draft sentence from the hypothetical sample, compare the reference content returned by retrieval with the original text of the declaration detail schedule, and observe whether the reference contains the subject, the period and the field name. If the reference is an empty array, record it as no hit; do not conclude that no such rule exists.
- Enter the same question through semantic search and full-text search separately, compare whether the two retrieved sets cover the declaration detail schedule number and the sales detail ledger amount field, and observe which path is more complete; use that observation to decide whether to switch to hybrid search.
- Enter the original question and the question-optimized version, compare whether the subject and period are consistent between them, and observe whether "July 2025" was rewritten into another period. If it was rewritten, record that the search terms need manual correction; do not treat this as evidence that the declaration data is wrong.