What the documents look like
Export tax rebate filing happens after goods leave the country, when tax vouchers are matched against the export record. The working materials are the export tax exemption and rebate filing form, the customs declaration form, the purchase invoice, and the foreign exchange receipt voucher. These come from the electronic tax bureau, the export rebate filing system, customs electronic messages, and the accounting books. A single shipment can carry a Chinese filing form alongside English-language trade documents, with product names, overseas buyer names, customs declaration numbers, and invoice numbers appearing in more than one language.
That mix shapes retrieval. Semantic search can match across languages, so a Chinese question may reach an English-language knowledge point. Full-text search suits declaration numbers, invoice numbers, model designations, and proper nouns, where an exact hit matters more than a near-synonym. The same batch of materials may hold a Chinese filing form and an English document side by side, so whether the search term and the stored content agree on subject, period, and field decides whether the recalled content can support later reconciliation. Filing runs monthly and per batch, which sets the rhythm for refreshing stored material.
Constraints at the retrieval step
The knowledge base search node can link one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be linked together. Search mode is a comparison decision, not a fixed setting: semantic, full-text, and hybrid retrieval each behave differently on the same question, and no single mode is assumed to fit every material type.
The citation limit is measured in tokens, not in a top-k item count. The minimum relevance value is a filter threshold and takes effect only under semantic retrieval or when result reranking is used. Question optimization performs coreference resolution and question expansion, and it adds one model call before the actual search. Citation content is returned as an array whose length may be zero, and the downstream path still runs even when nothing is retrieved. Retrieval and recall do not constitute authorization, and they do not isolate entities or permissions.
Hypothetical paper exercise: a trading company exports one consignment on the hypothetical date 10 March 2025. The hypothetical customs declaration number is HYP-EXP-20250310-001, the hypothetical purchase invoice number is HYP-INV-20250310-088, and the hypothetical foreign exchange receipt voucher number is HYP-FX-20250310-012. Three source sentences are available for comparison. "The customs declaration HYP-EXP-20250310-001 has a declaration date of 10 March 2025, and the exporter is Hypothetical Company A." "The purchase invoice HYP-INV-20250310-088 has an issue date of 8 March 2025, and the purchaser is Hypothetical Company A." "The foreign exchange receipt voucher HYP-FX-20250310-012 has a receipt date of 2 April 2025, and the payee is Hypothetical Company A." A draft sentence is under review: "The foreign exchange receipt voucher for this export consignment has been received, and the rebate filing materials are complete." The draft places receipt and completeness side by side, so each field must be checked against the source sentences.
Configuration and implementation checks
The table below lists settings or implementation checks tied to this page's documents and primary task. The hypothetical starting values apply only to this paper comparison and are not official defaults.
| Configuration item | Hypothetical starting point and reason | Comparison method |
|---|---|---|
| Linked knowledge bases | Link only rebate material bases sharing one vector model; check the vector model identifier first | If identifiers differ, exclude that base and do not run cross-model retrieval |
| Search mode | Full-text for number questions, semantic for natural-language questions | Observe recalled content and matched fields for the same test question under each mode |
| Citation limit | Hypothetical start of 3000 tokens, since the sample holds three short vouchers | Compare against smaller and larger values for content completeness |
| Minimum relevance | Hypothetical start of 0.5, after labelling relevant and irrelevant samples | Compare filtering effect only under semantic retrieval or reranking |
| Question optimization | Enable for follow-up questions with references, disable for explicit number questions | Compare subject in the original and optimized question for distortion |
| Result reranking | Enable when candidates are numerous, according to the setting | Compare ordering differences between rerank scores and vector scores |
Three common errors
- Treating "voucher received" as "materials complete." The source sentences state only the voucher number and receipt date; they do not state that the filing form, customs declaration, and purchase invoice agree on their fields. Input: the three source sentences and the draft sentence. Comparison: subject and period in each field. Observable result: the draft sentence goes beyond what the source supports, and the completeness claim remains unverified.
- Using minimum relevance to filter permissions or entities. Input: Hypothetical Company A material and Hypothetical Company B material in the same base. Comparison: the relevance value against the subject field. Observable result: relevance does not distinguish subjects, so the authorization boundary must be checked separately.
- Linking knowledge bases that use different vector models. Input: two bases with different vector model identifiers. Comparison: the model identifiers. Observable result: when identifiers differ, the bases cannot be linked, and no inference is drawn about missed recall or redundant results.
How to verify the setup
- Input the three source sentences and the draft sentence. Compare the receipt field against each field the completeness claim depends on. Observe whether the draft sentence is supported by the source. The source does not state that the filing form, customs declaration, and purchase invoice agree, so that part stays pending verification.
- Input the hypothetical customs declaration number HYP-EXP-20250310-001. Compare the recalled content under full-text search and under semantic search. Observe whether the number reaches the matching source sentence. If it does not, record the gap; a first-position hit is not guaranteed.
- Input an explicit paper empty result of []. Check how the downstream design handles a citation array of length zero. Observe whether the path still runs. This does not claim that a fictional rule search must return [].