What the documents look like
Export tax rebate filing happens after goods leave the country, and it matches tax vouchers to that export. The working materials are the export tax exemption and rebate filing form, the export customs declaration, purchase invoices, and foreign exchange receipt vouchers. These come from the electronic tax bureau, the export rebate filing system, customs electronic messages, and the accounting books. A product data and selection question has to answer which customs declaration, which purchase invoice, and which commodity code and rebate rate apply to a given shipment, and the answer must point back to those original vouchers. Materials update monthly or per shipment batch. Citation tracing requires that each answer can point to a specific material and version, because otherwise the rebate is delayed, the rebate is denied, or a tax risk appears.
For this page the task is product data and selection questions inside export tax rebate filing, and the feature under examination is citation source and tracing. A retrieved passage is candidate source text only. It does not authorize a filing, and it does not show that a record is eligible to be filed.
Constraints at the citation tracing step
The knowledge base search node selects one or more knowledge bases that use the same vector model, and it outputs cited content as an array whose length may be zero. Even when nothing is searched, that output path still runs through. The citation upper limit is controlled in tokens. The minimum relevance value takes effect only in semantic retrieval or when result reranking is used. Hybrid retrieval merges multiple result paths through RRF. Reranking requires a model and runs according to the setting. Same vector model is the precondition for joint association; for knowledge bases with different vector models, check the identifier first and exclude the mismatched one.
A hypothetical paper exercise sample: a foreign trade enterprise accountant checks the rebate materials for a batch of women's cotton knitted sweaters in August 2025. The source text supplied for checking reads: "Export customs declaration number A0001, export date 12 July 2025, commodity number 6110200090, quantity 1200 pieces." "Purchase invoice number B0002, invoice date 20 July 2025, specification consistent with the customs declaration, quantity 1200 pieces." The draft sentence to be judged reads: "This batch of goods has completed the rebate filing and has been approved." The source text states only customs and purchase information, and does not record filing submission, review, or approval status.
The optional paths in the overall retrieval flow do not all execute on every request. Image retrieval requires a corresponding image input and model capability. Reranking requires a reranking model and runs according to the setting. A relevance score is not a permission or entity filter. Retrieval is not authorization. The minimum relevance value does not exclude unrelated businesses or unauthorized records.
Configuration and implementation checks
| Configuration or check item | Value or action on this page | Basis and explanation |
|---|---|---|
| Associated knowledge bases | Select only rebate material knowledge bases with the same vector model | Joint association requires the same vector model |
| Search mode | Hypothetical exercise starting point: hybrid retrieval | Numbers and models suit full-text recall, semantic recall supplements natural language questions, compared with semantic retrieval alone |
| Citation upper limit | Hypothetical exercise starting point: 3000 tokens | Holds customs declaration and invoice passages, compared with 1500 tokens to observe truncation |
| Minimum relevance | Hypothetical exercise starting point: 0.3 | Takes effect only in semantic retrieval or reranking, compared with 0.5 to observe missed recall |
| Result reranking | Enable according to the setting when a reranking model is available | The rerank score is used for filtering; when not enabled it does not run |
| Citation output handling | Check whether the citation array is passed to the answer step | An empty array also runs through, and each item must be compared against the source text |
Three common errors
- Treating a retrieved customs declaration passage as evidence that a filing was submitted or approved. Check action: input the hypothetical sample source text and the draft sentence, compare whether the source text contains a filing status field, and when the source text contains only customs and purchase information, judge the draft sentence unsupported.
- Associating knowledge bases with different vector models together, then judging missed recall from one search result. Check action: check the vector model identifier of each knowledge base, compare whether the identifiers match, and when they do not match, exclude that knowledge base and do not run a cross-model retrieval comparison.
- Treating the minimum relevance value or a similarity score as a permission or enterprise isolation measure. Check action: input a record belonging to another entity together with the current question, compare whether a permission field appears, and when the score reflects relevance only, record that the permission judgment still lacks a basis.
How to verify the setup
- Input the customs declaration source text, the purchase invoice source text, and the draft sentence from the hypothetical sample. Compare whether the draft sentence contains an approval status not recorded in the source text. When the draft sentence contains "has been approved" and the source text has no such field, the draft sentence is judged unsupported.
- Input a paper empty citation array []. Compare whether the downstream answer step continues to run. When the path runs through and no error is reported, the empty-result handling matches the documented description.
- Input two knowledge base configurations whose vector model identifiers differ. Compare whether they are jointly associated. When they are excluded because the identifiers do not match, the same-model association requirement is confirmed, and the actual retrieval observation is recorded separately.
A pending request is not approval. An offered option is not execution. An unsigned termination is not a completed termination. Where the source text does not state a field, that field remains unknown.