What the documents look like
Export tax rebate filing happens after goods leave the country, and the work is matching tax vouchers to what was declared. The daily materials are the export tax exemption and rebate filing form, the export customs declaration, purchase invoices, and foreign exchange receipt vouchers. In the inventory procurement and supplier records step, purchase invoices and supplier information must be tied back to the customs declaration and the filing form in a traceable way. Data updates monthly or per shipment batch.
For citation tracing, the practical point is that every declared figure should point back to a source voucher. A retrieval hit is a candidate original text, and a retrieval result does not grant authorization, does not confirm supplier identity, and does not confirm rebate eligibility. The knowledge base search node can associate one or more knowledge bases that share the same vector model, and it outputs cited content as an array whose length may be zero. When the array is empty, that output path still runs through. The citation cap is measured in tokens. The minimum relevance value takes effect only in semantic retrieval or when result reranking is used; it is a filter threshold, not a sorting rule, and it does not perform permission or entity isolation. RRF is the fusion step for mixed retrieval results, not a switch that can be turned on. Thinking budget belongs to model reasoning strength configuration and has no relation to the citation token cap.
Constraints at the citation tracing step
This page addresses one task only: export tax rebate filing, inventory procurement and supplier records, citation source and tracing. Retrieval narrows candidates; it does not decide whether a filing may proceed, and it does not decide whether a supplier or a rebate claim is valid. Forms collect values, and business verification, approval, database writes, and external synchronization each require their own stated downstream implementation. A pending request is not an approval, an offered option is not execution, and an unsigned termination is not a completed termination.
Hypothetical drill sample. A finance staff member at a foreign trade enterprise is checking a draft filing for one March 2025 export shipment. The source materials provided for comparison read: "Purchase invoice number A-2025-0312, invoice date 10 March 2025, goods name consistent with the customs declaration, supplier is Company A." "Export customs declaration number B-2025-0087, export date 15 March 2025, domestic consignor is the foreign trade enterprise." The draft sentence to be judged reads: "The supplier records for this shipment have passed verification and the rebate may be filed directly." The phrase "have passed verification" has no corresponding record in the materials, so it is unsupported.
Two further boundaries follow from the same documents. First, whether a citation is adopted still depends on review against the effective document version; a current version and a historical version can both be retrieved, and version validity is a human check. Second, the retrieval scope must be confirmed against the selected knowledge bases and their document versions, and only knowledge bases with the same vector model may be associated together. This same-model requirement concerns the knowledge base association only; it does not extend to business registration sites or any other external system.
Configuration and implementation checks
The values below are hypothetical starting points for this drill, chosen so they can be compared against a counterpart value and adjusted. They are not official defaults.
| Configuration item | Hypothetical starting point for this drill | How to check and what it means |
|---|---|---|
| Associated knowledge bases | Only the inventory procurement and supplier records base | Check the vector model identifier of each base; exclude bases whose identifier differs |
| Search mode | Mixed retrieval | Compare against semantic retrieval alone and observe whether number-type terms are hit |
| Citation cap | Hypothetical starting point of 2000 tokens | Compare against total candidate source length and observe whether content is truncated |
| Minimum relevance | Hypothetical starting point of 0.3 | Compare against rerank scores and observe whether low-scoring content still enters the citations |
| Citation content output | Passed to the answer step | Check each answer sentence against the cited source text |
| Document version check | One current version and one historical version | Observe which version the citation points to; the effective version still requires human review |
The citation cap is expressed in tokens because chunk lengths vary widely across mixed knowledge bases, which makes a fixed top-k unstable. Mixed retrieval widens the candidate range and cannot be similarity-filtered directly, so reranking is usually needed before filtering by rerank score. The minimum relevance value is a filter threshold, not a final ordering rule; after question optimization, mixed retrieval, image retrieval, or reranking, results may be fused from several recall paths and are not necessarily ordered by a single vector similarity. Image retrieval applies only when an image input exists and the model supports the corresponding capability, and reranking applies only when a rerank model is available and enabled by the setting. The optional paths shown in the overall flow are not all executed on every request.
Three common errors
- Treating a retrieval hit as authorization. Input the draft sentence "have passed verification," compare it against the source materials, and observe that the materials state only the invoice and customs declaration details and contain no verification conclusion, so the sentence is unsupported. A retrieval result is not authorization, and it does not confirm supplier identity or rebate eligibility.
- Treating the minimum relevance value as a permission filter. Input a supplier records query, compare the returned citations, and observe that relevance affects ordering and filtering only; it does not exclude unauthorized records. A relevance score is not a category, entity, or permission filter.
- Treating an empty citation array as a workflow error. Input a paper-level empty array [] as the citation output, compare it against the downstream design, and observe that the path still runs through. Whether an error is raised depends on the downstream implementation, and the documents do not describe an automatic error. An empty reference output retains its documented condition.
How to verify the setup
- Input the draft sentence from the hypothetical sample together with the two source materials. Compare whether "have passed verification" has a corresponding record in the materials. The observable result is that the materials do not state that conclusion, so the sentence is judged unsupported. The evidence gap remains: no verification record exists in the supplied text.
- Input the vector model identifiers of the associated knowledge bases. Compare whether the identifiers match. The observable result is that bases with differing identifiers are excluded, and no cross-model retrieval is performed to compare missed recalls. Only the identifiers are checked and non-matching bases are excluded, with the actual observation recorded.
- Input a paper-level empty array [] as the citation output. Compare whether the downstream node accepts it. The observable result is that the path runs through; whether a business action is triggered depends on the downstream implementation, and the documents do not describe an automatic action. A pending state is not an approval, and a retrieval result is not an execution of any business step.