What the documents look like
Export tax rebate filing takes place after goods leave the country, and the work is matching tax vouchers to the export transaction. The working materials are the export tax exemption and rebate filing form, the export customs declaration, the purchase invoice, and the foreign exchange receipt voucher. These materials come from three separate sources: tax, customs, and finance. The filing cycle runs monthly or per shipment batch. The task of comparing contracts and transaction materials is to place contract terms, customs declaration information, invoice content, and receipt records under the same entity, the same period, and the same field, then judge whether the cited source can support the value entered on the filing form.
Because the sources are dispersed, citation tracing must land on a specific material and a specific version for each entry. Rebate delay, denial of the rebate, and tax risk are the consequences of error. The tracing step exists so that every cited data point can be returned to its original material.
Constraints at the citation tracing step
The knowledge base search node can select one or more knowledge bases that share the same vector model as the search scope. Its output is an array of citations, and the length may be zero. Even when nothing is retrieved, the output path still completes. The citation limit controls, in tokens, the maximum content cited per search. The minimum relevance is a 0–1 filter threshold, effective only under semantic retrieval or when result reranking is used. Hybrid retrieval merges multiple result paths through RRF, and reranking produces a 0–1 score. These mechanisms govern the recall and ordering of cited content. They do not perform permission judgment, and they do not perform entity or subject isolation. The requirement that retrieval libraries share the same vector model applies only between knowledge bases; it does not extend to business registration sites or other external systems.
Hypothetical exercise sample: a finance staff member at a foreign trade enterprise is organizing one export batch. The material text has three sentences. Sentence one is a customs declaration excerpt: "Consignor: Company A; date of departure: 12 March 2024; commodity name: men's cotton shirts; quantity: 5,000 pieces; total price: USD 45,000." Sentence two is a purchase invoice excerpt: "Buyer: Company A; invoice date: 28 February 2024; goods name: men's cotton shirts; quantity: 5,000 pieces; total amount including tax: RMB 310,000." Sentence three is a foreign exchange receipt excerpt: "Payee: Company A; receipt date: 20 April 2024; amount: USD 45,000; payer: Company B." The draft sentence to be judged is: "This export batch has received foreign exchange payment and the rebate filing may be submitted." The three materials span February to April, the entity is Company A throughout, the quantities match, and the receipt amount matches the customs total price.
Configuration and implementation checks
| Configuration or check item | Hypothetical starting point for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Select the knowledge bases holding the customs declaration, purchase invoice, and receipt voucher | First check the vector model identifier of each base; only matching identifiers may be associated together, and non-matching ones are excluded |
| Search mode | Hybrid retrieval | Contract numbers and customs declaration numbers are exact identifiers; semantic and full-text retrieval are used together and merged through RRF |
| Citation limit | Set to 3,000 tokens as the starting point for this exercise | This value is a starting point for this exercise, not an official default; adjust after comparing with the token volume of the three sample materials |
| Minimum relevance | Set to 0.5 as the starting point for this exercise | This value is a starting point for this exercise, not an official default; compare with the score distribution after reranking is enabled |
| Result reranking | Enabled according to the setting | The rerank score is used for filtering and is compared with the vector score obtained without reranking |
| Citation output handoff | Pass the citation array to the answering step | Check each value in the answer against the cited source text |
Three common errors
- Treating retrieval results as authorization evidence. Retrieving the customs declaration and invoice only shows that the materials were recalled within the search scope. It does not show that the rebate for this batch has been approved. The phrase "the rebate filing may be submitted" in the draft sentence requires a separate check of the filing authorization status; the retrieval output itself does not constitute approval.
- Treating evidence from different fields as a contradiction. In the sample, the customs total price is USD 45,000 and the purchase invoice is RMB 310,000. The currencies differ, so these are not mutually exclusive values in the same field. A contradiction requires clearly exclusive values under the same entity, the same period, and the same field.
- Treating the receipt fact as contract effectiveness or filing completion. The receipt voucher shows USD 45,000 received on 20 April 2024, which shows only that this payment occurred. Contract signing, cooperation reached, and contract effectiveness are three distinct facts. The sample provides no contract text, so signing and effectiveness status remain unsupported.
How to verify the setup
- Input the three sample material sentences and the draft sentence. Compare the phrase "has received foreign exchange payment" in the draft with the receipt date, amount, and payee fields on the receipt voucher. When payee, amount, and date all match, that clause is supported by the material. If either the payee or the amount differs, the clause is unsupported and the gap must be recorded.
- Input the phrase "the rebate filing may be submitted" from the draft sentence and compare it with the filing status field on the export tax exemption and rebate filing form. The sample provides no filing form, so this clause is pending verification. If a filing form is later provided and the status field shows accepted, the clause is supported; if it shows not submitted, the clause is unsupported. Retrieving the customs declaration does not change this judgment.
- Input the citation array and compare it with the source text cited in the answer. When the citation array length is zero, the downstream handoff design must handle an empty array and must not treat it as a workflow error. When the citation array is non-empty, check each value in the answer to confirm it can be traced to the same entity, the same period, and the same field in the corresponding material text. A value with no traceable source is recorded as unsupported, and the evidence gap is preserved.