What the documents look like
Export tax rebate filing happens after goods leave the country, and the work matches tax vouchers to that export. The core 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. Cross-border materials frequently carry Chinese and English side by side: declared product names, overseas buyer names, and receipt voucher summaries. During retrieval, fragments from different batches and different contracts can be pulled into one answer.
Citation tracing addresses a narrow question: can each voucher passage that appears in an answer be traced back to a specific document and a specific field. Materials update monthly and per batch. That frequency affects when reconciliation is performed; this page does not derive any synchronization mechanism from it.
Constraints at the citation tracing step
The knowledge base search node can select one or more knowledge bases that share the same vector model. Knowledge bases built on different vector models cannot be associated together. The vector model identifier must be checked first, and any base that does not match is excluded. Retrieval across different vector models is not performed in order to compare missed recalls.
The node outputs citations as an array, and the length can be zero. An empty array still passes through the output chain; it does not mean the workflow errored out and stopped. The citation content then goes to the answering step, where each answer statement is checked against the cited source text.
Similarity and the minimum relevance score act only on recall ordering and filtering. They do not perform permission isolation or entity isolation. Retrieval is not authorization. Hybrid retrieval merges multiple result paths through RRF, which is a result fusion process, not a switch that gets turned on. Reranking requires a model and must be enabled according to settings. Image retrieval requires image input and the corresponding model capability. Optional paths shown in the documentation flow diagram do not mean every request executes all of them.
Hypothetical drill sample: a finance staff member at a foreign trade enterprise is reconciling rebate materials in March 2026 for a batch exported in January 2026. The source text provided for checking is: "Export customs declaration number 2026XXXX01, export date 2026-01-12, domestic consignor Company A, overseas consignee Company B, commodity name stainless steel cookware, quantity 1200 pieces." A purchase invoice summary reads: "Invoice date 2026-01-05, buyer Company A, goods stainless steel cookware, quantity 1200 pieces, amount not stated." One draft sentence to be judged is: "This batch of stainless steel cookware has completed export, the purchase invoice and the customs declaration agree on quantity, and the rebate filing can be submitted." The judgment scope is limited to whether the subject, object, period, and fields are supported by the source text. "The rebate filing can be submitted" is a business review conclusion, outside what this page's materials can support.
Configuration and implementation checks
| Configuration or check item | Value or practice on this page | Basis and boundary |
|---|---|---|
| Associated knowledge bases | Select only rebate material bases whose vector model identifier is the same | Only the same vector model can be associated; check the identifier first and exclude non-matching bases |
| Search mode | Combine semantic retrieval and full-text retrieval, for product name and declaration number queries | Number-type queries suit full-text retrieval; natural language queries suit semantic retrieval |
| Citation limit | This hypothetical trial starts at 3000 tokens, compared with 1500 and 6000 for recall completeness | This value is a hypothetical starting point for this drill, not an official default |
| Minimum relevance | This hypothetical trial starts at 0.3, compared with 0.2 and 0.5 for filtered result count | Takes effect only under semantic retrieval or result reranking |
| Citation handoff | Pass the citation array to the answering step and compare each item against source text | An empty array also passes through; observe how the downstream presents it |
| Material version check | Test source pointing with the current version and a historical version | Whether a citation is adopted still requires review against the effective version |
Three common errors
- Treating a similarity score as permission or entity filtering. The input is a query containing an overseas buyer name; the comparison is whether materials of other subjects appear in the recalled results; the observable result is that the score level has no relation to subject ownership, and the material scope must be checked separately.
- Treating an empty citation array as retrieval failure. The input is a query unrelated to rebate materials; the comparison is whether the node output is an array of length zero; the observable result is that the output chain still passes through, and a failure branch must be defined separately in the downstream design.
- Treating "quantity agrees" in the draft sentence as an already-verified conclusion. The input is the customs declaration and purchase invoice text above; the comparison is the quantity field and subject field across the two materials; the observable result is that the quantity fields agree and the amount field is not stated, so the conclusion can only reach quantity agreement.
How to verify the setup
- Input the customs declaration text and purchase invoice summary above, compare whether the two judgments "has completed export" and "quantity agrees" in the draft sentence each have source text support, and observe the result: the former needs the declaration export date and commodity information, the latter needs the two quantity fields to agree, and the unstated amount remains a gap.
- Input an explicit paper citation array [], compare how the downstream answering step is designed to handle an empty array, and observe the result: the citation display is empty and the chain is not interrupted. No claim is made that a fictional rule's retrieval necessarily returns [].
- Input a retrieval request containing an overseas buyer name, compare whether the subject field in the recalled results is the same subject as the request, and observe the result: only the same subject, same period, and same field constitute an agreement or contradiction finding. Fields not stated, or evidence covering only a related stage, are recorded as pending verification, not as a negation.
This page covers only export tax rebate filing, cross-border materials and multilingual retrieval, and citation tracing. Whether materials, operations, or requests are in an authorized state at handover cannot be settled by retrieval or by a form.