What the documents look like
Producer export tax processes combine domestic-sales offset with export refund in one calculation. For the inventory procurement and supplier records task, four document families must line up item by item: the exemption, offset and refund declaration detail sheet; the export contract; the sales detail ledger; and the export customs declaration. Procurement warehouse receipts, supplier invoices and payment vouchers form the input side. Export customs declarations and the sales detail ledger form the outbound side. Citation tracing answers a narrow question for each inventory or supplier record: which material it came from, which version of that material, and whether that version was within its effective period.
The monthly update rhythm of the underlying data does not change the citation tracing step itself. What matters is that a retrieved fragment can be walked back to a specific document version, and that the answer text stays inside what that version states.
Constraints at the citation tracing step
The knowledge base search node can associate one or more knowledge bases, and those knowledge bases must share the same vector model. Knowledge bases built on different vector models cannot be associated together. The node outputs citations as an array, and that array may have length zero; an empty result still passes through the downstream chain.
The minimum relevance setting is a 0-1 filter threshold, and it takes effect only in semantic retrieval or when result reranking is used. It filters low-relevance fragments. It does not act as a permission filter, an entity filter, or a supplier isolation mechanism. The citation upper limit is measured in tokens, not in top k items. Full-text retrieval and semantic retrieval are merged through RRF, which is a fusion process rather than a switch that can be turned on. Image retrieval requires image input and matching model capability; reranking requires a rerank model and must be enabled in settings. Optional paths in the overall retrieval flow do not mean every request executes all of them.
Hypothetical paper exercise sample: a producer enterprise files a March 2025 exemption, offset and refund declaration. The inventory procurement material contains one VAT special invoice issued by Supplier A, goods name "polyester fabric", quantity 5,000 metres, invoice date 10 February 2025. One export customs declaration shows export date 25 February 2025, quantity 4,800 metres. Proposed draft sentence: "The procurement and export quantities of this batch of polyester fabric are consistent, and the supplier material can be confirmed as the complete procurement source for this batch of exemption, offset and refund." The question to decide: does the source text support this draft sentence?
Configuration and implementation checks
| Configuration item | Setting or check relevant to this page | Correspondence with the sample |
|---|---|---|
| Associated knowledge bases | Associate only knowledge bases that share the same vector model | Compare the vector model identifiers of the inventory procurement base and the supplier material base |
| Search mode | Semantic retrieval, full-text retrieval or hybrid retrieval | Use full-text retrieval for exact fields such as invoice number and customs declaration number |
| Citation upper limit | Controls the volume of cited content in tokens | Hypothetical trial starting point 2,000 tokens; the sample material is short, so compare against 4,000 tokens to observe citation completeness |
| Minimum relevance | A 0-1 threshold, effective only in semantic retrieval or reranking | Used only to filter low-relevance fragments, not to exclude suppliers and not as a permission control |
| Citation content output | An array, length may be zero | Paper input [] to check whether the downstream design accepts an empty result |
| Reranking | Requires a rerank model and must be enabled in settings | May be enabled when the sample produces many candidates, to observe ordering changes |
The 2,000-token starting value is a hypothetical trial value chosen because the sample materials are short; it is compared against 4,000 tokens to observe whether more cited content changes completeness. It is not an official default.
Three common errors
- Treating minimum relevance as a supplier or permission filter. Input: retrieval over the sample invoice and customs declaration. Comparison: the relevance score of each fragment against the threshold. Observable result: fragments scoring below the threshold are filtered out, while other batch records from the same supplier are not excluded merely because their relevance score is low. Relevance does not perform entity isolation.
- Forcing association of knowledge bases built on different vector models. Input: the vector model identifiers of the inventory procurement base and the supplier material base. Comparison: whether the two identifiers match. Observable result: where the identifiers differ, the bases cannot be associated together; record the actual observation and do not infer missed recalls or redundancy.
- Treating a pending review status as granted authorization. Input: the sample draft sentence and the source material text. Comparison: the quantity field 5,000 metres against 4,800 metres. Observable result: the two values differ, so the draft sentence claim of consistent quantities is not supported. Where the material states no authorization or effective clause, that gap remains open for verification.
How to verify the setup
- Input the sample invoice text "quantity 5,000 metres" and the customs declaration text "quantity 4,800 metres". Compare the values of the same field. Observable result: the two values are mutually exclusive, and the draft sentence claim of consistent quantities is judged unsupported.
- Input a paper citation array []. Compare whether the downstream node still executes. Observable result: when the citation length is zero, the chain passes through, and an empty array is not judged as an error that terminates the flow.
- Input the vector model identifiers of the inventory procurement base and the supplier material base. Compare whether the identifiers are the same. Observable result: only matching identifiers permit association; where identifiers differ, exclude that base and record it, without running a cross-model retrieval to compare missed recalls.
Citation tracing answers only which material and which version a given record came from. Retrieval and form collection gather values; they do not constitute authorization, approval or a database write. Whether supplier material is complete, and whether procurement was authorized, require downstream business verification and human confirmation.