What the documents look like
Producer export tax refunds combine domestic-sale tax offsets with export refunds in one calculation. That structure shapes citation tracing: one line in the refund declaration detail schedule can draw on domestic sales, export sales, customs declarations, and the financial ledger at the same time, so any quoted figure has to be traceable to a specific document. The documents handled day to day are the refund declaration detail schedule, export contracts, the sales detail ledger, and export customs declarations. Cross-border materials often carry Chinese, English, or other language fields, and a search result that returns a passage without naming its source removes the basis for comparing free-on-board price differences or for grouping self-produced exports. Data updates monthly, and this page does not derive batch synchronization or real-time triggers from that cadence.
Constraints at the citation tracing step
The knowledge base search node outputs reference content as an array, and that array may have length zero. When nothing is found, the output path still runs. Citation tracing therefore separates two situations: when the reference array is non-empty, check each entry against its source text; when the array is empty, check how the downstream design receives an empty array. The node can link one or more knowledge bases that share the same vector model. Bases with different vector models cannot be linked together. At the check step, confirm the knowledge base identifiers first and exclude those that do not qualify; do not run cross-model retrieval to compare missed recalls. Hybrid retrieval merges multiple result paths through RRF, and RRF is a fusion process, not a switch that can be turned on. Minimum relevance takes effect only in semantic retrieval or when result reranking is used. It is a filter threshold and does not perform permission or entity isolation, and retrieval itself does not constitute authorization.
Hypothetical exercise sample: a check of March 2026 refund materials for an export producer. The source texts are "Export contract HT-2026-031, counterparty overseas buyer A, goods are self-produced hardware parts, free-on-board terms, signed 5 March 2026"; "Export customs declaration 20260312001, declared 12 March 2026, export date 10 March 2026, transaction terms FOB"; and "Sales detail ledger, March 2026 voucher 18, summary is export sales, amount agrees with the customs declaration." The draft sentence to be judged is "This export completed the refund declaration in March 2026 and the refund was received." The materials needed for the judgment must sit in a target retrieval scope the checker has access rights to, or in explicit manual input. Where they are missing, record the gap.
Configuration and implementation checks
| Configuration or check item | Hypothetical starting point for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Linked knowledge bases | Link only two bases sharing one vector model: the refund declaration detail schedule base, and the export contracts and customs declarations base | Official guidance requires the same vector model before bases can be linked together; confirm identifiers first and exclude those that do not qualify |
| Search mode | Hybrid retrieval | Contract numbers and customs declaration numbers are exact identifiers, so semantic and full-text retrieval run together and merge through RRF |
| Reference limit | Hypothetical starting point 3000 tokens for this exercise | Official guidance controls the reference limit in tokens; the reason is that schedule and contract passages differ widely in length; compare against a smaller value of 1500 tokens and observe how the number of referenced entries changes |
| Minimum relevance | Hypothetical starting point 0.4 for this exercise | Official guidance states the value takes effect only in semantic retrieval or result reranking; compare against 0.2 and observe entries added or removed |
| Result reranking | Enabled, with a rerank model already configured | Official guidance states reranking produces a 0-1 score that can be used for filtering; compare ordering before and after enabling |
| Reference content handoff | Check whether the retrieval output is passed to the answering step | The official output is a reference array whose length may be zero; check each answer sentence against the referenced source text |
Three common errors
- Treating an empty reference array as a workflow error. The input is the search node output []. The comparison object is the node documentation stating the length may be zero. The observable result is that the output path still runs. Judging the empty array a failure is an incorrect conclusion.
- Treating minimum relevance as a permission or entity filter. The input is a low-scoring entry inside the same retrieval scope. The comparison object is the official statement that the value is a filter threshold, not a final ordering rule. The observable result is that the low-scoring entry is filtered while unauthorized records are not excluded by that setting.
- Treating retrieval or a form as authorized execution. The input is the draft sentence "completed the refund declaration and the refund was received." The comparison object is that the source texts record only contract signing, customs declaration, and bookkeeping. The observable result is that the source texts do not state declaration completion or receipt of a refund, so the draft sentence is unsupported.
How to verify the setup
- Input the contract text, the customs declaration text, and the ledger text from the sample. Compare the export date and transaction terms fields across the three for the same entity and the same period. The observable result is that the customs declaration export date of 10 March 2026 corresponds to the ledger summary, while the contract does not state an export date and is recorded as pending verification.
- Input the draft sentence "This export completed the refund declaration in March 2026 and the refund was received." Compare its field coverage against the three source texts. The observable result is that the source texts contain no record of declaration completion or refund receipt, so the sentence is judged unsupported and the evidence gap is retained.
- Input an explicit manual value of [] as the retrieval output. Compare it against the design description of the downstream answering step. The observable result is that the downstream step receives the empty reference without interruption. This paper input checks the design; it does not state that a fictional-rule retrieval necessarily returns [].