What the documents look like
Producer export tax credit and refund combines domestic sales tax offset with export refund in one calculation. This combined calculation shapes what store service and internal training must trace when citing sources. The materials to reconcile are the credit-and-refund declaration detail table, the export contract, the sales subsidiary ledger and the customs export declaration. These four document types carry different field definitions, so a figure drawn from one does not automatically align with a figure drawn from another. Data updates monthly, and any conclusion cited in a service or training setting must first be tied to the same declaration period across the materials being compared.
Errors in the tax amount can lead to a reduced refund or a corrected tax filing. For that reason, a cited source must be traceable to a specific material and a specific field. The declaration detail table records the period and the export sales total; the export contract records signing date and agreed delivery period; the sales subsidiary ledger records export revenue for the month. None of these fields, on its own, records an approval outcome.
Constraints at the citation tracing step
The knowledge base search node selects one or more knowledge bases that share the same vector model and outputs a citation array whose length may be zero. An empty array means the chain ran through without obtaining citation content; it does not show that retrieval failed. Citation results must be passed to the answering step, where each answer is checked line by line against the cited original text. The minimum relevance value is a filtering threshold and does not carry permission or entity isolation functions. Retrieval scope is set by the selected knowledge bases and is unrelated to the relevance number.
Hypothetical drill sample: a store service staff member for an export producer prepares an internal training script. The source text is: "The credit-and-refund declaration detail table (hypothetical number MX-2024-07) states the period July 2024, with the export sales column totalling 1.2 million yuan; the export contract (hypothetical number HT-2024-015) is signed on 20 June 2024 and sets delivery in August 2024; the sales subsidiary ledger records export revenue of 1.2 million yuan for July 2024." The draft sentence is: "This enterprise's July 2024 export business has completed refund approval." The question to decide: can the cited materials support the phrase "has completed refund approval"?
Configuration and implementation checks
| Configuration item | Hypothetical starting value for this drill | How to check |
|---|---|---|
| Linked knowledge bases | Select knowledge bases containing the credit-and-refund declaration detail table and the export contract | Compare the vector model identifier of each base and exclude any that do not match |
| Citation limit | Hypothetical starting value 3000 tokens, chosen because a single declaration detail table carries many fields | Compare against the actual tokens of one material and observe whether truncation occurs |
| Minimum relevance | Hypothetical starting value 0.4, chosen to retain exact hits for number-based queries | Compare with full-text retrieval hits and observe how many records remain after filtering |
| Search mode | Hybrid retrieval, since number and model queries depend on keyword hits | Observe semantic and full-text recall results separately for the same question |
| Question optimisation | Not enabled for a single clear question | Compare recalled content with the setting on and off |
| Citation output | Receives the array, length may be zero | Use a paper input of an empty array to check the downstream design |
Three common errors
- Treating a retrieval hit as business authorisation. A returned citation only shows that text was matched; it does not show that a refund was approved or that a contract took effect. Check: input the hypothetical source text above, compare the citation content against the phrase "has completed refund approval", and observe whether an approval outcome field appears in the citation.
- Treating minimum relevance as a permission filter. The relevance value does not distinguish enterprises, stores or authorisation scope. Check: within the selected knowledge bases, input a hypothetical cross-entity question, compare the returned citation against the knowledge base scope, and observe whether content outside that scope appears.
- Treating an empty citation array as a retrieval failure. When the output length is zero, the chain still runs through. Check: use a paper input of an empty array to check the downstream node design and observe whether the downstream handles it as an empty citation, without inferring that retrieval necessarily returns empty.
How to verify the setup
- Verification one: input the hypothetical source text, compare the draft sentence "This enterprise's July 2024 export business has completed refund approval" against the fields the source contains, and observe whether an approval status field appears in the citation content. The source contains only the declaration period, the sales amount and the contract date, and states no approval conclusion. The sentence is unsupported and the evidence gap must be recorded.
- Verification two: input a question containing the declaration detail table number, compare the recall results of semantic retrieval and full-text retrieval, and observe whether the number is hit exactly. If full-text retrieval hits while semantic retrieval does not, the number query depends on the keyword path.
- Verification three: use a paper input of an empty array to check the downstream receiving design and observe whether the downstream node handles a citation length of zero. This check does not claim that retrieval necessarily returns an empty array; it only examines the downstream handling logic for an empty citation.