What the documents look like
Producer export tax processes combine the offset of domestic-sale tax against export refund in a single calculation. That structure means comparison work has to cover both the export side and the domestic-sale side at once. The four document types handled day to day are the exemption, offset and refund declaration detail sheet, the export contract, the sales detail ledger and the export customs declaration. Each carries a different basis: the declaration sheet carries the filing basis, the contract carries the agreed basis, the ledger carries the book basis and the customs declaration carries the customs basis. Data sources include the electronic tax bureau, ERP sales, customs electronic information and the financial books, and these sources differ in how fields are named and how periods are divided.
Comparison work concentrates on three stages: collecting self-produced exports, calculating the exemption, offset and refund, and handling free-on-board price differences. The task on this page is retrieval and recall from a knowledge base: taking a search query, returning candidate fragments from material already stored, and handing those candidates to a person who checks whether the contract and the transaction material agree on subject, object, status and effective scope. Retrieved results are candidate content only. They do not constitute authorization, approval or business confirmation.
Constraints at the retrieval step
A knowledge base search node can associate one or more knowledge bases that use the same vector model. Bases using different vector models cannot be associated together. When model identifiers differ, check the identifiers first, exclude the bases that do not match, and record what was actually observed. Do not run a cross-model search in order to compare recall differences.
A single search does not run every path every time. Semantic search, full-text search and hybrid search are chosen according to the test question. Image description retrieval or image vector retrieval requires an image in the input and a model that supports it. Result reranking requires a rerank model and must be enabled in the settings. Hybrid search merges multiple channels through RRF; RRF is a fusion process, not a switch that can be turned on. Quoted content is output as an array and its length can be zero, so the output path still completes when nothing is found.
Minimum relevance takes effect only under semantic search or when result reranking is used. It is a filtering threshold and does not perform permission or entity isolation. The reference limit is measured in tokens. Question optimization performs coreference resolution and question expansion, and adds one model call before retrieval.
Hypothetical paper exercise. A producer export enterprise library contains one export contract fragment reading: "Party A is a domestic production enterprise, Party B is an overseas buyer; the goods are self-produced A-type motors, quantity 1,200 units; the trade term is FOB, shipment period March 2026; the contract amount follows the invoice confirmed by both parties." The library also contains one sales detail ledger fragment reading: "March 2026, A-type motors exported, 1,200 units, booked amount consistent with the export invoice, corresponding customs declaration number hypothetical BG-2026-03-001." The draft sentence to be judged reads: "The shipment period for the 1,200 A-type motors under this contract and the export date stated on the customs declaration fall in the same period, so the contract and the customs declaration can be confirmed as compared and consistent." Judging this draft sentence requires checking whether the contract shipment period, the customs declaration export date and the ledger period share the same subject, the same period and the same field.
Configuration and implementation checks
| Configuration item | Starting point for this hypothetical exercise | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the exemption, offset and refund library that uses the same vector model | Compare the vector model identifiers of each base; exclude any that do not match |
| Search mode | Test numbering-type questions with full-text search first, then test natural-language questions with semantic search | Observe the difference in recalled fragments between the two modes using the same set of questions |
| Reference limit | Hypothetical starting point of 3,000 tokens | Test whether long and short contract and ledger fragments are returned complete, and compare against results at a shorter limit |
| Minimum relevance | Hypothetical starting point of 0.3 | Label relevant and irrelevant samples first, then compare against results at adjacent thresholds |
| Question optimization | Enabled for follow-up questions containing pronouns | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Enabled when candidate fragments are numerous | Compare candidate order before and after enabling, and observe where relevant fragments move |
These values are starting points for this paper exercise. They are not official defaults and were not executed against any real record.
Three common errors
- Treating a retrieval hit as authorization confirmation. The input is the draft sentence to be judged. If retrieval returns the contract fragment and the customs declaration fragment, the comparison result shows only that the fragments were recalled. Contract signing, customs declaration filing and refund approval are separate facts, and retrieval produces no authorizing effect.
- Treating an unstated field as a contradiction. The input is the contract fragment and the sales detail ledger fragment. The contract states a shipment period and the ledger states a booking period; these are different fields. Only an explicit mutually exclusive value in the same field constitutes a contradiction. A field that is not stated remains to be verified.
- Forcing bases with different vector models into a shared association. The input is the vector model identifiers of several bases. If the identifiers do not match, exclude the non-matching base and record the gap. Do not infer what recall would look like after a shared association.
How to verify the setup
- Input the contract text, the sales detail ledger text and the draft sentence from the hypothetical exercise. Compare whether the contract shipment period, the ledger period and the customs declaration export date share the same subject, the same period and the same field. If the field bases agree and the values are compatible, the draft sentence is supported. If the customs declaration export date is not stated, the result is pending verification.
- Input the phrase "the contract and the customs declaration can be confirmed as compared and consistent" from the draft sentence. Compare it against what the recalled fragments actually state. If the recalled fragments do not contain the customs declaration export date, this confirming phrase is not supported, and the evidence gap is recorded.
- Input the list of vector model identifiers for the associated bases. Compare whether the identifiers match. If any do not match, exclude that base and record the actual observation. If all match, continue testing by search mode and observe whether recalled fragments cover the contract, the ledger and the customs declaration. Any part not covered remains a gap.